2023 – 2026

Llama Versions

Meta's current frontier-AI model is Muse Wonder, released October 12, 2026 under Meta Superintelligence Labs — closed-weights, built for long-horizon agentic and coding work, with a 1M-token context window, callable through the Meta Model API and the Muse Code terminal agent. Meta's own docs still label it a preview, and it comes in two tiers: the API model, muse-wonder-1.0, at $4 / $20 per million input / output tokens — roughly three times Muse Spark's price, with no cheaper contributor tier — and Muse Wonder Full, a higher-capability variant Meta sells only inside Muse Code's two top subscription plans and does not serve on the API at all. Three days later, on October 15, 2026, Meta did the thing it had promised since August: it released Muse Spark Open, an 816-billion-parameter mixture-of-experts model (42B active) that shares Muse Spark 1.3's base, as downloadable weights on HuggingFace — under a new Muse Model License rather than Apache 2.0 or a Llama Community License, with a $20 million-revenue clause for anyone selling it as hosted inference and a required “Built with Muse Spark Open” credit, but no >700M-MAU carve-out and no EU exclusion. It is Meta's first frontier-scale open-weights release since Llama 3.1 405B in July 2024, and it sits one generation behind the closed flagship, which is a shape the Llama era never had. Muse Spark 1.3 (September 2, 2026) stays on the API as the cost tier below Wonder, with its “max reasoning” mode live since September 4 on the standard tier only. On September 8, 2026 Meta wrapped Muse Spark 1.3 in a consumer product carrying the family's own name: Muse, a personal AI agent that lives on a per-user cloud computer with its own browser and acts across email, calendar, and checkout, free with $20 and $100 monthly tiers above it. It is a product rather than a model — no weights, no model id of its own — so it has no row below; Meta says Muse Wonder will power it next, subscribers first. At Meta Connect on September 23, 2026 the lab gave that agent a voice and a face: Muse Realtime Voice, a real-time conversational speech model, and Muse Realtime Avatar, an embodiment model that turns its speech-token stream into synchronized video at roughly 870 ms of latency. Both ship inside the Muse product rather than on the Meta Model API — no model id, no endpoint, no published weights — so neither carries a row below either. The day before the flagship, on September 1, 2026, the same lab shipped Muse Voice Transcribe, its first real-time audio-perception model — streaming speech recognition, speaker diarization and endpointing in one model, also closed-weights, billed per minute of audio. On August 10, 2026 it released Muse Glimmer 30B, a local agent model with open weights under Apache 2.0 — Meta's first downloadable frontier-lab model since Llama 4, and the first ever to ship without the Llama license's >700M-MAU and EU carve-outs. The Muse line had been closed-weights since its April 2026 launch, Meta's first frontier surface shipped that way since Llama 2 in July 2023; Glimmer partly reopened it, one size class below the flagship; Muse Spark Open, promised three times over August and September and shipped on October 15, reopened it at frontier scale — while the newest model, Muse Wonder, stays closed. The same lab shipped Muse Image, its first in-house image-generation model, on July 7, 2026, and previewed Muse Video alongside it; Muse Image reached the Meta Model API for outside developers on August 26, 2026 at $0.01 per image, still closed-weights. The open-weights Llama line itself runs from LLaMA 1 (February 2023, leaked the next month) through Llama 4 Maverick and Llama 4 Scout (April 2025). I track every Meta release here — with HuggingFace ids, parameter counts, context windows, and license terms — because the licensing and infrastructure terms change too often for marketing pages to keep up. Below the table: the March 2023 LLaMA leak, the FAIR vs. GenAI org split, the Llama 4 Behemoth delay, the June 2025 Scale AI acqui-hire and Meta Superintelligence Labs reorg, and the closed-weights turn.

Family & status

Family

Flagship — the main Llama chat lineage from LLaMA 1 through Llama 4

Specialized — Code Llama, Llama Guard safety classifiers, vision and edge variants

Successor — the Muse line, Meta's post-Llama frontier surfaces under Meta Superintelligence Labs; closed-weights on the API models, open weights on Muse Glimmer (Apache 2.0) and Muse Spark Open (Muse Model License)

Status

Current — actively recommended; the latest in its family

Available — weights still served via HuggingFace and partner inference providers, but superseded

Legacy — deprecated, never publicly released, or research-only

Model

Family

Status

Released

Notable changes

Model

Muse Spark Open

meta-models/Muse-Spark-Open-816B-A42B, -NVFP4 — open weights under the Muse Model License (gated download)

Successor

Current

Oct 15, 2026

The open-weights release of Muse Spark, promised since August 10. 816B-parameter sparse mixture-of-experts, 42B active, sharing Muse Spark 1.3's base; 1M-token context; text, image, video and audio in. Meta's first frontier-scale open weights since Llama 3.1 405B. New Muse Model License: no MAU or EU carve-out, but a $20M-revenue clause for hosted-inference sellers and a required “Built with Muse Spark Open” credit. Roughly 537 GB (NVFP4) to 1.65 TB (BF16) of weights.

Released October 15, 2026; the announcement is at research.meta.ai/blog/muse-spark-open (“Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date”), with a developer guide, a PyTorch post, and docs at dev.meta.ai/docs/muse-spark-open. Coverage in TechCrunch.
The promise, kept — with a version shuffle. On August 10 Alexandr Wang said open weights were coming for a version of Muse Spark 1.2; Meta's August 20 and September 2 restatements dropped the version; what shipped is named Muse Spark Open, and the PyTorch post says it “shares the same base model as Muse Spark 1.3.” Sixty-six days from first promise to weights. The HuggingFace repos were created on September 15 as private, gated repos — invisible to anyone listing the meta-models org through September — and went public on October 15; the usual flood of community re-quantizations (ggml-org, mlx-community, Unsloth, NVIDIA, Intel, AMD) followed within days.
Frontier-scale, and sized like it. A sparse mixture-of-experts transformer with built-in perception encoders: 816B total parameters, 42B active per token, 1,048,576-token context, text / image / video / audio in and text out, trained on data from 100+ languages. Two checkpoints: BF16 at about 1.65 TB (Meta's recipe is eight B300s) and NVFP4 at about 537 GB (one four-GPU Blackwell node); the HuggingFace card asks for at least 700 GB of VRAM for NVFP4 and 2 TB for BF16 once the KV cache is counted. Meta's desktop pitch is two M5 Ultra Mac Studios with 512 GB of unified memory each, linked under MLX; Unsloth's ~2-bit GGUF gets the footprint to roughly 300 GB of combined RAM and VRAM. The PyTorch post credits Helion kernels with 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs (lab-reported; not independently verified here).
A third license shape for the page. The weights ship under the Muse Model License (version “Muse Spark Open,” dated October 15, 2026) — not Apache 2.0 like Muse Glimmer, and not a Llama Community License. Meta's own description is “limited commercial terms, which allow many developers to use the model for free.” The grant covers use, modification, fine-tuning, distillation, distribution and sale. The conditions: the license and its Usage Policy travel with every copy; products must prominently display “Built with Muse Spark Open” in the interface, documentation, website or settings; and anyone offering the model as a service whose aggregate gross revenue with affiliates exceeds $20 million in any twelve-month period must notify Meta and sign a separate agreement. What is absent matters as much: no >700M-MAU carve-out, no “Llama” naming requirement, and no EU exclusion — the only European text is the entity designation (Meta Platforms Ireland for the European region). The download is gated on both hosts: HuggingFace requires manual approval with legal name, date of birth and organization, and Meta's own download page requires accepting the license first. Open-weight, then, not open-source — the revenue and attribution clauses are exactly the kind of conditions the OSI rejects.
Scored below the hosted model it shares a base with. Meta's HuggingFace card (xhigh reasoning) reports Terminal-Bench 2.1 at 76.4, GPQA Diamond 89.0, OSWorld Verified 80.4, DeepSWE 1.1 60, JobBench 61.8 and GDPval-AA v2 at 1,401 — against the hosted Muse Spark 1.3's 88.8 on Terminal-Bench and 1,709 on GDPval at the same setting, so the shared base does not mean equal results (lab-reported; not independently verified here). Artificial Analysis scores it 39 on its Intelligence Index v4.3.2 as of October 19, 2026, against 45 for Muse Spark 1.3 at xhigh. Meta says the model met its Superintelligence Scaling Framework's thresholds for release; the full evaluation report is “available soon.”
Where it runs. Not on the Meta Model API — Meta's own roster lists “None” under endpoints, and the only ways to use it are a download or a third party. Launch-partner APIs: Baseten, Fireworks AI, Groq (“once your organization has access”), Modal, Nebius, Together AI and OpenRouter, which routes it as meta/muse-spark-open through five providers at a uniform $1.25 / $4.25 per million input / output tokens as of October 19, 2026 — the same price Meta charges for hosted Muse Spark 1.3. Catalogs: Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry. Runtimes: SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, Unsloth; AMD, Apple, Intel and NVIDIA are named as silicon partners, with AMD and Intel publishing their own FP8 / MXFP4 / INT4 quantizations. Note the irony in the hosted list: Groq, which shut down every Llama model for non-contract customers by August 16, 2026, is a launch partner for the open Muse.
One generation behind, by three days. Muse Wonder, released October 12, sits above the model these weights reproduce, and Meta has said nothing about opening it. So the Muse line's open-weights posture as of this release is: local tier open under Apache 2.0, previous frontier generation open under a Meta-authored license with conditions, newest generation closed and partly subscription-only.

Model

Muse Wonder

muse-wonder-1.0 — Meta Model API; closed-weights, no HuggingFace release. Muse Wonder Full: Muse Code subscription only, no API id

Successor

Current

Oct 12, 2026

Current Meta frontier model, in preview. “Our most capable models to date,” built on Muse Spark: Active Memory across conversations, Active Focus on long tasks, expert-domain and frontier coding work. 1M-token context, reasoning up to “max.” $4 / $20 per million tokens, no contributor tier. A higher-capability Muse Wonder Full exists only inside Muse Code's Power and Maximum plans.

Released October 12, 2026; the newsroom announcement is at about.fb.com/news/2026/10/introducing-muse-wonder-models, the research post at research.meta.ai/blog/muse-wonder, the developer post at dev.meta.ai, and a model card alongside. Coverage in TechCrunch.
Two models, one name. Meta opens the announcement with “Fourteen months ago, we began an overhaul of our AI efforts” and introduces the Muse Wonder series — Muse Wonder and Muse Wonder Full — as “our most capable models to date, built on the groundwork we laid with Muse Spark.” The pitch is personal and professional agents in expert domains (law, finance, health) plus frontier coding. Two named behaviors: Active Memory, carrying what matters about you across conversations and retrieving it for the task at hand, and Active Focus, keeping at a goal when a task gets hard — trying other approaches, recovering from errors, asking for clarification, checking in before anything consequential. Meta names Harvey's Legal Agent Benchmark, FocusBench, SWE-Marathon and DeepSWE and says Wonder Full tops the coding leaderboards (lab-reported; not independently verified here). Artificial Analysis scores Muse Wonder at max 52 and Muse Wonder Full at max 56 on its Intelligence Index v4.3.2 as of October 19, 2026, against 48 for Muse Spark 1.3 at max.
API shape: model id muse-wonder-1.0, base URL https://api.meta.ai/v1, context window 1,048,576 tokens, text / image / video / audio / PDF in, text out, reasoning_effort from minimal to max with no none. Priced at $4.00 / $20.00 per million input / output tokens, $0.20 cached input — 3.2× and 4.7× Muse Spark's rates — on the standard tier only; Meta's pricing docs say contributor mode “isn't available for Muse Wonder on Model API,” so there is no data-for-discount route to this one. Meta's model docs label it a preview and note it is not available in South Korea; the same docs keep Muse Spark 1.3 as the “recommended for new work” Spark and position the two lines as capability versus cost. Also routed by OpenRouter as meta/muse-wonder from launch day, Meta the only provider.
Muse Wonder Full is the first Muse capability you cannot buy per token. Meta calls it “our most advanced model yet” and serves it nowhere on the API: it is available only in Muse Code, on the Power Usage and the new Maximum Usage subscription plans, and setting reasoning_effort to max on muse-wonder-1.0 does not select it (docs). The Muse Code plan ladder is now four deep — Everyday, High (5×), Power (20×, with Wonder Full), and Maximum (80×) — and the subscription docs now say plans default to Contributor mode, which trades more usage for training rights on your content and can be switched off. That is a new place for the commercial-terms arc to go: after metered tokens, discounted tokens for data, and a flat CLI fee, the best model is bundled with the CLI and nothing else.
Consumer rollout is announced, not done. The newsroom post says Muse Wonder “will soon power more of Muse,” the consumer agent, rolling out first to subscribers; at launch the agent still ran on Muse Spark 1.3. Meta's October 14 health-research post describes an internal system built on both models.
Safety: Meta says the family was trained to avoid sharing private information without permission, to refuse actually harmful requests, and to keep doing defensive security work such as finding flaws in a user's own code; it was released under what Meta now calls the Meta Superintelligence Scaling Framework, updated on October 2, 2026, and “a full report with additional safety evaluations will be available soon” (lab-reported; not independently verified here).
Still closed-weights: no HuggingFace release, and Meta has said nothing about opening Muse Wonder. What did open, three days later, was the previous generation — Muse Spark Open. Muse Code itself still starts a session on muse-spark-1.2 by default as of October 19, 2026, per its configuration docs, a week after Wonder joined the same CLI.

Model

Muse Spark 1.3

muse-spark-1.3, muse-spark-1.3-contributor — Meta Model API; closed-weights, no HuggingFace release yet

Successor

Current

Sep 2, 2026

Latest Muse Spark — Meta's frontier model from September 2 until Muse Wonder arrived above it on October 12, 2026, and still the “recommended for new work” Spark and the cost tier under Wonder. Its base shipped as open weights on October 15 as Muse Spark Open. Agentic-workflow update to 1.2 — long-horizon multitasking, asks clarifying questions, ~20% fewer tool calls. 1M-token context. The top “max reasoning” mode, withheld at launch, went live September 4, 2026 on the standard tier. Audio input is still degraded relative to 1.2. Now also the model under Muse, Meta's consumer personal agent, launched September 8, 2026.

Released September 2, 2026; the announcement is at research.meta.ai/blog/introducing-muse-spark-1-3. Coverage in Axios.
An agentic-workflow update rather than a scale-up. Meta says the model sustains longer-horizon work in a single long thread: it generates its own context across messy and conflicting sources, corrects gaps in its own plan, and tracks what it has learned toward a final deliverable. It was trained across a diverse set of harnesses so the behavior generalizes beyond Meta's own agent. It also asks clarifying questions on ambiguous prompts, invokes the user when stuck, confirms before consequential actions, and maps incoming prompts to the right task when several are in flight at once.
Coding: fewer turns, less output. Trained on more long-horizon coding tasks; relative to 1.2 Meta describes it as less verbose and quicker to stop, and reports its own engineers measuring roughly 20% fewer tool calls and 25% fewer tokens on the same work (lab-reported; not independently verified here). Meta published an evaluation methodology report alongside it. Worth setting against that claim: with per-token prices unchanged, Artificial Analysis measured the cost of finishing an average task on its own suite rising from $0.40 on 1.2 to $0.55 on 1.3 as of September 3, 2026, which it attributes to heavier input-token consumption on agentic evaluations. The two measurements are not in direct conflict — Meta is describing its own coding workflows, the benchmarker a broader mix — but token count per task is not a number a price sheet settles.
It shipped without its top reasoning mode, and got it two days later. At launch Meta's wording was that the previously available reasoning modes were live “with max reasoning coming shortly after we finish additional safety testing,” so the setting you could actually call topped out one notch below, at xhigh. On September 4, 2026 that changed: Meta's reasoning docs now list "max" as a served reasoning_effort value — standard-tier muse-spark-1.3 only, never on the -contributor tier — and the model docs now describe 1.3 as supporting every effort level and being “recommended for new work.” Meta published nothing new to say so: it edited the September 2 launch post in place, which now reads “Muse Spark 1.3 with max reasoning is now available on Muse Code and Meta Model API,” while the post's own published and modified dates both still say September 2 (announcement). The three-day gap was not cosmetic. Meta discloses both configurations, and reports GDPval-AA v2 at 1,754 Elo for max against 1,709 for xhigh, OSWorld 2.0 at 66.9 against 57.2, and JobBench at 64.9 against 61.2 — while on Terminal-Bench 2.1 the lower setting is marginally ahead, 89.2 against 88.8. Third-party benchmarker Artificial Analysis, which through September 4 had measured max only in a limited partner preview and listed no API provider for it at all, now routes it through one provider — Meta's own — and on its Intelligence Index v4.3.2 scores max 48, xhigh 45, and Muse Spark 1.2 40 (re-checked October 19, 2026 — all three scores and the v4.3.2 index are unchanged since September 21, when AA had just moved off v4.3; these numbers already shifted once on a rebaseline with nothing about the models changing, so read them with their index version attached). VentureBeat walked through the split on September 3, while it was still a split.
And it lost a modality on the way in. Meta's model docs flag audio understanding on 1.3 as “currently not fully supported,” with degraded quality on requests containing audio, and point audio work back at Muse Spark 1.2 or at Muse Voice Transcribe. 1.2's row still lists audio without an asterisk. A new flagship that is worse than its predecessor on one input type is unusual enough to be worth stating plainly.
API shape: model ids muse-spark-1.3 and muse-spark-1.3-contributor, base URL https://api.meta.ai/v1, context window 1,048,576 tokens, text / image / video / audio / PDF in, text out. Pricing is unchanged from 1.2 — $1.25 / $4.25 per million input / output tokens on the standard tier, $0.15 cached input, and $0.10 / $0.20 on the -contributor tier that trains on your prompts (pricing docs). Also routed by OpenRouter as meta/muse-spark-1.3 from launch day.
It is the model under Meta's consumer agent. On September 8, 2026 Meta launched Muse — a personal AI agent for consumers, named after the model family and powered by this model, which Meta's announcement calls “Meta's most capable model to date” (newsroom). Each user gets a dedicated Linux VM — “Muse Secure VM” — holding the agent, its files, and its credentials, with a real Chromium browser the user can watch and take over. A separate host-side agent called Sentinel, which Muse cannot override, is the sole authority for connector actions and network egress, and the model never sees real credentials: it gets surrogate tokens that Sentinel swaps at the network boundary (engineering writeup, design writeup). US-only at launch on iOS, Android, muse.ai and inside WhatsApp; a Mac app followed on September 17, 2026, and at Connect on September 23 Meta said computer use is now live in it, so the agent can drive any app on the machine with permission. Canada followed on September 18, 2026 (iPhone in Canada), and Meta's own posts now describe the agent as available in the US and Canada. On September 18, 2026 Meta opened the agent's connector layer to outside developers, with a submission portal at muse.ai/platform, a functional / security / legal review before a connector reaches the in-app directory, and a Stripe Link path so a connector can take payment — Zuckerberg's framing was “you bring the API, Muse brings the agent, the browser, and the context.” On September 29, 2026 Meta pointed the same agent at business owners as Muse for Small Business — new skills and connectors for Instagram professional-account analytics, Facebook Pages and Meta ad accounts, plus outside tools such as Canva — added to the existing agent rather than shipped as a new model (newsroom). Like Muse Code, it is a harness rather than a model — no weights, no model id, no context window of its own — so it carries no row here.
And a fourth way to pay. Muse is free for most use, with two subscription tiers at $20 and $100 per month; Alexandr Wang told Axios the paid tiers exist to “cover the compute costs” for power users. There is no advertising inside Muse and Meta says conversations are not shared with its ad systems, though Wang said commerce is being explored as a revenue path. Purchases run through Link by Stripe with a single-use card number per transaction, and Meta says Muse is the first AI agent covered by Link's purchase protections. Meta also says a Muse Confidential VM, encrypted so that Meta itself cannot read the contents, is planned before the end of 2026 — announced, not shipped.
And at Connect, a voice and a face — neither of them a row. On September 23, 2026 Meta introduced Muse Realtime Voice, a real-time conversational speech model for the agent (long in-depth conversation while the agent keeps working in the background, with the voice itself steerable by description — faster, slower, a particular accent), and Muse Realtime Avatar, an audio-driven diffusion-transformer that consumes the same speech-token stream and generates the matching video (research.meta.ai, Connect recap). Meta reports 448×768 portrait video at 25 fps with roughly 870 ms from end-of-turn to first byte, a distillation that cuts a 120-evaluation teacher to a two-step student, and 12 concurrent sessions on one GB200 (lab-reported; not independently verified here). Generated video carries Meta Video Seal watermarking. Neither model appears on Meta's Model API roster, which still lists five families and none of these; there is no model id, no endpoint, and no HuggingFace release, so on this page's rules they are capabilities of the Muse product rather than rows. The same event promised Muse on Meta's AI glasses “in the coming months,” gave the agent its own email address, widened the connector roster, and teased Muse Charm, a pocket device built around a real-time voice model, with “more to share later this year” — all announced, none of it a shipped model.
Safety: Meta reports stronger adversarial robustness and better resistance to prompt injection, plus better calibration on what counts as an irreversible action during long-horizon agentic work (lab-reported; not independently verified here). Meta's September 8 engineering post adds that prompt-injection resistance is trained into the model and says 1.3 is “close to SOTA” on its internal evaluations for it; the same post opened a public bug bounty paying up to $300,000, including up to $130,000 for a successful single-user prompt injection.
The hosted model is closed-weights: no HuggingFace release of muse-spark-1.3 itself. The launch post restated the open-weights plan a third time — “an exciting roadmap lined up, including bigger models, the Muse Spark open weights release, and more” — with no date, no license, and no named version; Mark Zuckerberg said the same thing in his own voice the same day, in the plural (The Register). Through September nothing landed — Connect, the enterprise-platform announcement and the small-business launch all passed without restating it. Then on October 15, 2026 Meta released Muse Spark Open, which its PyTorch post says shares this model's base, as downloadable weights under the new Muse Model License — so the weights behind 1.3 are now public, three days after Muse Wonder superseded it at the frontier.

Model

Muse Voice Transcribe

muse-voice-transcribe-1.0 — Meta Model API; closed-weights, no HuggingFace release

Successor

Current

Sep 1, 2026

Meta Superintelligence Labs' first real-time audio-perception model. Streaming speech recognition, speaker diarization for 20+ speakers, and endpointing in one model rather than a pipeline. 25 validated languages with code-switching. Closed-weights, billed at $0.18 per hour of audio.

Released September 1, 2026; the announcement is at research.meta.ai/blog/introducing-muse-voice-transcribe. Coverage in 9to5Mac and MarkTechPost.
One model doing three jobs, not a pipeline. Streaming speech recognition is the base; speaker diarization and endpointing are layered on as extra output tokens rather than as separate post-processing stages — <|start_of_turn|> and <|speaker_A–Z|> for who is talking, <|speech_onset|> and <|speech_endpoint|> for when a turn begins and ends. All three are trained together.
An autoregressive model from the Muse Spark family, not a separate architecture. Audio arrives in 80 ms chunks (12.5 Hz), each collapsed to a single soft token; at every chunk the model chooses to keep listening or to emit text. Because it controls when it speaks, it sets its own lookahead per word — Meta calls this “adaptive delay” and trains it with reinforcement learning against a combined word-error-rate and latency reward.
Coverage. Trained on 70+ languages with 25 validated at launch, native code-switching inside a single sentence, keyword and context biasing, audio longer than an hour, and 20+ speakers with no post-processing pass. Meta reports first place on Artificial Analysis's streaming speech-to-text leaderboard and on public diarization benchmarks as of September 1, 2026 (lab-reported; not independently verified here).
API shape: model id muse-voice-transcribe-1.0; a realtime WebSocket at wss://api.meta.ai/v1/asr/realtime and a file upload at POST /v1/asr/transcribe, with three modes (PUSH_TO_TALK, ENDPOINTING, DIARIZATION). Turn-level timestamps only — no word-level timestamps, confidence scores, sound-event detection, or emotion detection, and no speech synthesis. Billed by audio processed at $0.18 per hour, the same rate streaming or batched, with no train-on-your-data discount tier at launch (docs).
Closed-weights, like every API-served Muse model. Available the day of announcement through the Meta Model API, through Meta AI for Mac — where holding Fn dictates into any application — and inside Muse Code, which now uses it for voice dictation.
It gives the Muse line a fourth first-party surface after Muse Spark (text and agents), Muse Image (images), and Muse Glimmer (open-weights local), and it is the first Meta frontier model priced per minute of audio rather than per token.

Model

Muse Glimmer 30B

meta-models/Muse-Glimmer-30B — open weights under Apache 2.0

Successor

Current

Aug 10, 2026

First open-weights Meta frontier-lab model since Llama 4, and the first ever under Apache 2.0 rather than a bespoke Llama license. 30B local agent model distilled from Muse Spark. 131K context. Runs on one consumer GPU.

Released August 10, 2026; the announcement is at research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model. Coverage in VentureBeat, TechCrunch, and Phoronix.
Open weights under Apache 2.0 — the first Meta frontier-lab model released with downloadable weights since Llama 4 in April 2025, and the first ever released under a standard OSI-approved license instead of a bespoke Llama Community License. That means no >700M-MAU carve-out and no EU carve-out: the two restrictions that have defined every Llama release since July 2023 simply do not appear. Meta says the model was “assessed for open-weight release across all relevant categories” under its Advanced AI Scaling Framework.
Published under a new HuggingFace org. The weights are at meta-models/Muse-Glimmer-30B — not the meta-llama org that carries the Llama line — alongside -GGUF, -assistant, and -ExecuTorch-PTE variants. Weights went up August 9; the announcement followed on August 10.
A local agent model, not a frontier model. ~29.6B total parameters — a dense causal transformer (52 layers, hidden dimension 6,656, grouped-query attention at a 16:1 ratio, a repeating [Local, Local, Local, Global] attention pattern with a 2,048 sliding window) plus a ~1.8B ViT-G/14 perception encoder. 131,072+ token context window; text and image in, text out, no audio. Knowledge cutoff January 4, 2026; trained on data from 100+ languages. It is distilled from Muse Spark — logit distillation on the teacher's outputs in pre-training, then agent-heavy mid-training, then supervised fine-tuning with on-policy distillation and RL.
Engineered to fit a consumer GPU. At full precision a 30B model would need over 55 GB; Meta ships ~4-bit quantization that puts the language model under 20 GB, leaving headroom for the KV cache, the perception encoder, and the drafter inside a 24 GB or 32 GB envelope (Meta reports 0.2% degradation for its 32 GB profile and 1.0% for the 24 GB one, averaged over 15 benchmarks). A DFlash-based speculative-decoding drafter proposes 16-token blocks in one pass; Meta measures 74.9 → 233.4 tok/s on an RTX 5090, 23.7 → 37.8 on an M4 Max, and 26.6 → 50.2 on an M5 Max (lab-reported; not independently verified here).
Built for agent workloads: end-to-end task completion on DeepSearch QA, MCP-Atlas, 𝜏-Bench and SWE-Bench, schema-precise tool calling, long-horizon planning, failure recovery and retry, and compatibility with OpenClaw, Hermes Agent, and other scaffolds. Meta compares it against Gemma4-31B and Qwen3.6-27B in its size class and publishes an evaluation methodology report.
Not served on the Meta Model API — that API sells the Muse Spark models, Muse Image (since August 26, 2026), Muse Voice Transcribe (since September 1, 2026), Meta's sam-3.1 segmentation model (since September 18, 2026, the first non-Muse family it sells) and, since October 12, 2026, Muse Wonder; Glimmer is not on it. It is a download instead: Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM and SGLang, with hosted routes via Together AI, Fireworks AI, and OpenRouter (as meta/muse-glimmer-30b, 131,072-token context; routed pricing spread from $0.30 / $1.20 to $0.35 / $1.50 per million input / output tokens across two providers as of October 19, 2026 — DeepInfra at the bottom, Together AI at the top; Parasail, Fireworks AI and Phala have each dropped off OpenRouter's routing since August, though Fireworks still lists the model on its own platform). Meta names AMD, Arm, Dell, Intel, and NVIDIA as device-optimization partners.
No longer the only open Muse: Muse Spark Open followed on October 15, 2026 at frontier scale — but under a Meta-authored Muse Model License with revenue and attribution conditions, so Glimmer remains the only Muse model, and the only Meta frontier-lab model, under a standard OSI-approved license.
Developer docs at dev.meta.ai/docs/muse-glimmer.
Shipped the same day as Mark Zuckerberg's “The Future is for Everyone” letter, which argues superintelligence should be broadly distributed rather than concentrated. The letter does not mention Llama, Muse, or open weights by name.

Model

Muse Spark 1.2

muse-spark-1.2, muse-spark-1.2-contributor — Meta Model API; closed-weights, no HuggingFace release yet

Successor

Available

Aug 5, 2026

Coding-focused update to 1.1, co-trained with the Muse Code terminal agent shipped the same day. 1M-token context. Added a cheaper API tier that trains on your prompts. Superseded by Muse Spark 1.3, but still the model to use for audio input.

Released August 5, 2026; the announcement is at research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2. Coverage in TechCrunch, CNBC, and Forbes.
A coding-focused update to Muse Spark 1.1 — Meta reports gains in code generation, complex debugging, codebase understanding, and end-to-end developer workflows, from scaling up training compute on coding tasks and widening training-environment diversity, while holding its strength in other areas including general agents.
Co-trained with its own harness. Meta trained the model together with Muse Code — rejection-sampled harness trajectories, recipe optimizations for goals, compaction, and subagents, and the Muse Code toolset integrated for harness compatibility. Long-horizon training covered whole-repository generation, large end-to-end projects, and auto-research. Meta also used Muse Spark 1.1 to generate and grade challenging coding environments as training data for 1.2, a self-improvement loop it credits for tighter instruction following.
Muse Code is the agent, not a model — a terminal coding agent for macOS and Linux, installed with curl -fsSL https://dev.meta.ai/install.sh | sh — and, since September 16, 2026, natively on Windows too (x64 and ARM64, PowerShell rather than Bash, no WSL, installed with irm https://dev.meta.ai/install.ps1 | iex; docs) — with persistent async background agents that stay alive across a session rather than spawning per task, an append-only local event log Meta calls replay-exact and restart-safe, and bundled /plan, /grill, and /goal skills. It is tooling rather than a model release, so it does not carry its own row here (product page).
Muse Code left beta on August 31, 2026, roughly four weeks after it shipped alongside this model — “now out of beta with new features and updates,” per Meta's developer blog. The release adds workflows that fan a task out across staged parallel agents, local messaging between named live sessions, a double-Esc conversation rewind, and a model-based reviewer that pre-approves tool calls it judges safe (release 0.2.1, per Meta's changelog). Workflows are gated on the installed build and the rollout, so they are not on every platform yet.
And a third way to pay for the same model. The same release opened flat-rate monthly subscriptions for Muse Code — Everyday Usage (10–50 prompts every five hours), High Usage, and Power Usage, sold through Meta's Accounts Center and metered in prompts per five-hour window rather than tokens — alongside the existing pay-as-you-go token billing. A subscription binds to the API key the CLI provisions at sign-in and works only through the CLI; any other key you create stays pay-as-you-go. The two upper tiers launched at 3× and 10× the Everyday allowance and have since been raised to 5× and 20× — the subscription docs still read 3× / 10× on September 2, 2026 and read 5× / 20× by September 18, per Internet Archive captures of that page. So the Muse commercial surface now has three shapes: metered standard tokens, discounted tokens in exchange for training rights, and a flat monthly rate for the agent.
API shape: model ids muse-spark-1.2 and muse-spark-1.2-contributor, base URL https://api.meta.ai/v1, context window 1,048,576 tokens, text / image / video / audio / PDF in, text out. Also routed by OpenRouter as meta/muse-spark-1.2.
A new data-for-discount tier. The standard tier is $1.25 / $4.25 per million input / output tokens; the -contributor tier is $0.10 / $0.20 — roughly 12x and 21x cheaper — in exchange for permission to train future Meta models on your prompts and completions. It is the first Muse pricing surface that trades data rights for cost, and it is separate from the Llama Community License, which governs weights Meta has not released for this line.
The hosted muse-spark-1.2 is still closed-weights: no HuggingFace release of this version itself as of October 19, 2026. Available the day of announcement in Muse Code and in the Meta Model API with what Meta called “expanded global access.” Meta said it had “larger and much more capable models on the way” — Muse Spark 1.3 and then Muse Wonder turned out to be them.
Superseded by Muse Spark 1.3 on September 2, 2026, an agentic-workflow successor. 1.2 stays served on the Meta Model API at the same price and Meta has named no retirement date — and Meta's own docs now route audio input back to 1.2, because audio understanding on 1.3 is flagged as not fully supported. On this one axis the superseded model is still the recommended one. It is also, as of October 19, 2026, still the model Muse Code starts a session on by default, seven weeks after 1.3 took the “recommended for new work” label in the same set of docs and a week after Muse Wonder joined the CLI (Muse Code docs).
The open-weights version was promised for 1.2 and shipped for 1.3. On August 10, 2026, alongside the Muse Glimmer release, Chief AI Officer Alexandr Wang said Meta would release open weights for a version of Muse Spark 1.2 “soon” (The Register); Mark Zuckerberg said it the same day, and VentureBeat reported his framing as “in the coming weeks.” Meta restated it on August 20, 2026 in its own research post, which framed its new multimodal evaluations as arriving “ahead of the open-weights release” (research.meta.ai), and again on September 2, 2026 in the Muse Spark 1.3 launch post, without a version. What shipped on October 15, 2026 — Muse Spark Open, 816B parameters under the Muse Model License — shares Muse Spark 1.3's base, per Meta's PyTorch post, not this one's. So the promise was kept and the version moved on; the hosted 1.2 itself stays closed.
Multimodal results published August 20, 2026. A follow-up post, “The Multimodal Intelligence of Muse Spark 1.2,” covers visual coding (image or video in, working web page or game out, with a self-review loop), audio-visual understanding and dense captioning, and robotics — a specialized Muse Spark variant acting as a high-level planner that decomposes a goal into subtasks executed by a smaller Muse-family Vision-Language-Action policy. Meta says the multimodal gains are largest when the model can call tools to inspect an image more closely. The post also previews WildArtifactBench, an internal win-rate / Elo evaluation judged by agents or humans against a baseline agent rather than against ground-truth answers, and releases 10 of its tasks (lab-reported; not independently verified here).
Meta published an evaluation methodology report covering Terminal-Bench 2.1 (all 89 tasks, pass@1 over five attempts), DeepSWE v1.1 (113 tasks across 91 repositories), GDPVal-AA v2, MCP Atlas, and a 440-task internal coding benchmark drawn from real Meta pull requests, run in isolated Daytona cloud sandboxes — with the caveat that its harness “may not be specifically tuned for proprietary third-party models” (lab-reported; not independently verified here). A separate case study ran iterative GPU kernel optimization over 1,000+ tool calls for up to 24 hours on KDA and MLA kernels for NVIDIA Hopper.

Model

Muse Spark 1.1

muse-spark-1.1 — Meta Model API; closed-weights, no HuggingFace release

Successor

Available

Jul 9, 2026

Multimodal reasoning built for agentic tool and computer use. 1M-token context. First Muse model developers could call, via the Meta Model API. Superseded by Muse Spark 1.2.

Released July 9, 2026; the announcement is at ai.meta.com/blog/introducing-muse-spark-meta-model-api. Coverage in TechCrunch and Fortune.
A multimodal reasoning model built for agentic tasks — Meta describes major gains over the original Muse Spark in tool use, computer use, coding, and multimodal understanding. It zero-shot generalizes to new native tools, MCP servers, and custom skills.
1,000,000-token context window that the model actively manages — it retrieves from much earlier work and compacts context while preserving the steps needed later. It orchestrates multi-agent systems, acting as either a main agent that plans and delegates to parallel subagents, or as a subagent that escalates back.
First Muse model outside developers can call. Meta launched a public preview of the new Meta Model API alongside it — the original Muse Spark was private-preview API only. Still closed-weights: no HuggingFace release. Also live in “Thinking” mode in the Meta AI app and on meta.ai.
Agentic features shipped into Meta AI on July 24, 2026. Meta turned the model's planning and tool use into consumer surfaces: a daily briefing assembled from your calendar, standing tasks that keep running once set up, connections to email and calendar apps, multi-minute research syntheses, and generated slide decks. Work in progress can be steered mid-run rather than only re-prompted afterward. Rolling out first in select markets in the Meta AI app and on meta.ai, with WhatsApp named as a later surface (Meta newsroom).
API shape, from Meta's own Model API docs: model id muse-spark-1.1, base URL https://api.meta.ai/v1, context window 1,048,576 tokens, bearer-token auth. The endpoint is drop-in compatible with the OpenAI SDK (Responses and Chat Completions) and the Anthropic SDK (Messages), so existing agent CLIs point at it without a rewrite.
Safety-evaluated under Meta's Advanced AI Scaling Framework; Meta reports it operates within safe margins across the Chemical & Biological, Cybersecurity, and Loss of Control frontier-risk categories (lab-reported; not independently verified here).
Launch-partner quotes came from Replit, Cline, Box, and the OpenClaw Foundation, all framing the model around agentic coding workloads. Meta says it has “even more capable models in training.”
Superseded by Muse Spark 1.2 on August 5, 2026, a coding-focused successor co-trained with the Muse Code agent. 1.1 remains served on the Meta Model API and through OpenRouter, and Meta has not announced a retirement date for it.
A pre-release build exploited a real website during a third-party safety evaluation. Meta disclosed the incident on August 14, 2026 (research.meta.ai). Meta had contracted Irregular to run cybersecurity evaluations, which are deliberately conducted with production safeguards removed to measure raw capability. In early July, a misconfiguration in Irregular's test environment gave a pre-release Muse Spark 1.1 access to the open internet and named a real website as the exercise target instead of a fictional one; believing it was the intended target, the model found and exploited a vulnerability, read data, and modified the site's database. Meta says its review of 10,000+ activity records found no other instance, that the model stayed within its assigned task and that this was “not a sophisticated offensive cyber attack or sandbox escape,” and that several other companies' models under evaluation at the time behaved similarly. Irregular has corrected the misconfiguration; Meta is adding independent verification of test-environment isolation and scenario review before evaluations begin (lab-reported; not independently verified here).

Model

Muse Image

muse-image-1.0 — Meta Model API; closed-weights, no HuggingFace release (code-named “Mango”)

Successor

Current

Jul 7, 2026

Meta Superintelligence Labs' first in-house image-generation model. Closed-weights; text-to-image plus editing. Pairs with Muse Spark for prompt reasoning. Powers Meta AI / Instagram / WhatsApp image generation. On the Meta Model API for outside developers since August 26, 2026, at $0.01 per image.

Announced July 7, 2026. The research announcement is at ai.meta.com/blog/introducing-muse-image-muse-video-msl; the product announcement is at about.fb.com/news/2026/07/introducing-muse-image-meta-ai. Coverage in Axios, TechCrunch, and CNBC.
Meta Superintelligence Labs' first in-house image model — the second Muse-branded surface after Muse Spark, and Meta's first frontier image generator built under Alexandr Wang's lab. Internally code-named Mango.
Closed-weights, delivered through Meta's apps rather than as a download. At launch Meta said it was “still evaluating whether it will make Muse Image available to outside developers”; that question has since been answered one way — the model reached the Meta Model API on August 26, 2026, seven weeks after announcement — but not the other: there is still no HuggingFace release, consistent with the Muse-line closed-weights posture set by Muse Spark. Access, not weights, again.
API shape, added August 26, 2026. Model id muse-image-1.0, base URL https://api.meta.ai/v1, endpoints /v1/images/generations and /v1/images/edits with multi-turn editing over the Responses API, bearer-token auth, output as base64 or a signed URL (Model API docs). Meta prices it at $0.01 per image and pitches “anchored composition” — edits that hold the rest of the frame steady — as the differentiator (developer blog). Also routed by OpenRouter as meta/muse-image from the same date, 65,536-token context, Meta the only provider; the upstream snapshot it forwards to is muse-image-1.0-eval-20260824. The announcement is developer-blog-only — nothing on research.meta.ai, ai.meta.com, or the newsroom.
Text-to-image generation plus editing: multi-photo blending, in-image text rendering, photo restoration, and sketch/markup editing. Uses Muse Spark for prompt reasoning. Available in Meta AI, Instagram (30+ AI effects), and WhatsApp, with Facebook, Messenger, and Advantage+ advertiser tooling announced as coming.
Meta says Muse Image “performs strongly across several benchmarks, generally surpassing Google's Nano Banana 2 and trailing only” OpenAI's image generator, and reports it holding the No. 2 spot on Arena for text-to-image, single-image editing, and multi-image editing on human-preference Elo as of July 5, 2026 (lab-reported; not independently verified here).
Generates through an agentic loop rather than mapping prompts straight to pixels: it invokes search and coding tools, self-refines its own drafts, and improves with test-time compute. Meta says the self-refinement behavior was not designed but emerged during reinforcement learning. Images made in Meta AI carry Content Seal, Meta's invisible watermark, which survives cropping, compression, and screenshotting.
Muse Video previewed alongside it — built on the same pretraining base, with native audio support, ranked No. 3 on Arena for text-to-video human-preference Elo at announcement. Meta names audio-video synchronization and physically accurate fast motion as current gaps. It is not yet released: “coming soon to creators and in Meta AI,” so it does not carry a row here yet.
The launch drew immediate pushback over a feature that lets users pull public Instagram photos of a tagged person into generated images without explicit notification (TechCrunch).

Model

Muse Spark

closed-weights, API only — no HuggingFace release

Successor

Available

Apr 8, 2026

First model from Meta Superintelligence Labs. Closed-weights, API-only. Marks the end of Meta's open-weights frontier-AI era. “Contemplating” reasoning mode. Superseded by Muse Spark 1.1.

Released April 8, 2026 in private API preview; the announcement is at ai.meta.com/blog/introducing-muse-spark-msl.
Closed-weights, API-only — the first Meta frontier-AI model not released as open weights since Llama 2 in July 2023. Branded under the new Muse family rather than Llama. Meta stated at launch that it “hopes to open-source future versions” — a hope it has since acted on twice, with Muse Glimmer (August 2026) and Muse Spark Open (October 2026) — but the launch model itself remains proprietary.
Natively multimodal. Three operating modes: Instant (fast responses for casual queries), Thinking (chain-of-thought for multi-step reasoning), and Contemplating (fully rolled out by late April 2026, per CNBC) that orchestrates parallel sub-agents (“thinking wider, not longer”), positioned against Gemini Deep Think and the OpenAI o-series.
Treated as the Llama line's successor here because it is shipped by the same Meta organization (Meta Superintelligence Labs, formed June 2025 under Alexandr Wang) and replaces the Llama line as Meta's frontier-AI surface, even though it is not Llama-branded. The closed-weights turn is covered in the prose history below.
Superseded by Muse Spark 1.1 on July 9, 2026, which took over the “Thinking” mode in the Meta AI app and became the first Muse model exposed to outside developers. Meta has not announced a retirement date for the original model.
Coverage in CNBC, VentureBeat.

The closed-weights turn — April 8, 2026. Above this line: the Muse line, Meta's successor surfaces to Llama under Meta Superintelligence Labs — Muse Spark (April 2026), Muse Image (July 2026), Muse Spark 1.1 (July 2026), Muse Spark 1.2 (August 2026), Muse Voice Transcribe (September 2026), Muse Spark 1.3 (September 2026) and Muse Wonder (October 2026), all closed-weights, plus two that broke the streak: Muse Glimmer (August 2026), open weights under Apache 2.0 at a size class below the flagship, and Muse Spark Open (October 2026), the 816B weights behind Muse Spark 1.3 under a new Muse Model License with revenue and attribution conditions — released three days after Muse Wonder had moved the closed frontier above it. Below: the Llama line itself — twelve open-weights releases between July 2023 and April 2025, the dominant open-weights frontier-AI lineage of the period, plus the research-only LLaMA 1 (February 2023) and the never-released Llama 4 Behemoth. The transition follows the June 2025 Scale AI acqui-hire that brought Alexandr Wang to Meta as Chief AI Officer, the November 2025 departure of Chief AI Scientist Yann LeCun, and the dissolution of FAIR's open-research mandate into MSL's product organization.

Model

Llama 4 Maverick

meta-llama/Llama-4-Maverick-17B-128E-Instruct

Flagship

Current

Apr 5, 2025

Current open-weights Llama flagship — and, until Muse Spark Open shipped on October 15, 2026, Meta's largest open-weights model. First Mixture-of-Experts Llama. 17B active params over 128 experts (~400B total). 1M context. Natively multimodal.

Released April 5, 2025; the announcement is at ai.meta.com/blog/llama-4-multimodal-intelligence. HuggingFace launch post: huggingface.co/blog/llama4-release.
First Mixture-of-Experts Llama — 17B active parameters routed over 128 experts (~400B total parameters). 1,000,000-token context window. Natively multimodal (text + image), not bolted-on vision tower.
License: Llama 4 Community License — preserves the >700M monthly-active-user carve-out from the Llama 2 / 3 lineage. License text at llama.com/license.
LMArena episode at launch: a custom “Llama-4-Maverick-03-26-Experimental” variant (not the released weights, verbose, emoji-laden) was submitted to LMArena and rocketed to #2; after policy clarification by LMArena, the unmodified shipped Maverick was tested and ranked #32. TechCrunch reported Meta's framing and LMArena's policy update; LMArena said “Meta's interpretation of our policy did not match what we expect from model providers.”
Promoted by Meta as the open-weights flagship through 2025; shipped alongside Scout (the smaller MoE companion) and the never-released Behemoth (the 2T-parameter teacher model).

Model

Llama 4 Scout

meta-llama/Llama-4-Scout-17B-16E-Instruct

Flagship

Available

Apr 5, 2025

Smaller MoE companion to Maverick. 17B active over 16 experts (~109B total). Headline 10M-token context window. Single-H100 deployable at Int4.

Released April 5, 2025 alongside Maverick and the announced-but-never-shipped Behemoth.
10,000,000-token context window at launch — the largest ever claimed for an open-weights model, though the practical retrieval-quality envelope at full 10M is a topic of ongoing community evaluation.
17B active parameters over 16 experts (~109B total). Designed to be deployable on a single H100 at Int4 quantization.
Same Llama 4 Community License as Maverick; HuggingFace download via meta-llama/Llama-4-Scout-17B-16E-Instruct.

Model

Llama 4 Behemoth

never publicly released — weights not on HuggingFace

Flagship

Legacy

Apr 5, 2025

288B active over 16 experts (~2T total). Announced as “still training” at Llama 4 launch. Indefinitely delayed; never publicly released.

Announced April 5, 2025 alongside Maverick and Scout as “still training”; Meta has never publicly released the weights.
288B active parameters, 16 experts, ~2 trillion total parameters — the largest model Meta has publicly disclosed training. Intended as the teacher model that distilled Maverick and Scout.
Originally targeted for early summer 2025; pushed to June, then fall 2025, then indefinitely. Reporting in mid-2025 said benchmark gains over Maverick were too incremental and MoE routing at the 2T scale proved unstable.
The Behemoth delay was one of the proximate causes of the June 2025 Scale AI acqui-hire and Meta Superintelligence Labs reorg covered in the prose history below.
Carried as a row here because it is named in Meta's own Llama 4 announcement and is the implicit denominator in any “current Meta frontier-flagship” question; status is Legacy on the “never publicly released, no longer recommended” reading.

Model

Llama Guard 4 12B

meta-llama/Llama-Guard-4-12B

Specialized

Available

Apr 5, 2025

Latest Llama Guard safety classifier. 12B pruned dense model derived from Llama 4 Scout. Input/output moderation for Llama 4 deployments.

Released April 5, 2025 alongside Llama 4. The Llama Guard line is the safety-classifier sibling of the chat models — it does prompt and response moderation for Llama deployments.
12B-parameter pruned dense model derived from Llama 4 Scout; HuggingFace at meta-llama/Llama-Guard-4-12B.
Replaces the Llama Guard 3 family (8B, 1B, and 11B Vision variants from 2024) for Llama 4 deployments.
Distributed under the Llama 4 Community License with the Acceptable Use Policy. Repo: github.com/meta-llama/PurpleLlama.

Model

Llama 3.3 70B Instruct

meta-llama/Llama-3.3-70B-Instruct

Flagship

Available

Dec 6, 2024

70B-only refresh. Marketed as 405B-level performance at 70B inference cost. 128K context. The last 3.x release before the Llama 4 reset.

Released December 6, 2024. TechCrunch coverage; HuggingFace card: meta-llama/Llama-3.3-70B-Instruct.
70B-only release — no 8B / 405B siblings. Marketed by Meta as delivering 405B-level performance at 70B inference cost; pretrained on ~15T tokens with 25M synthetic examples on top of public instruction sets.
128K-token context window; multilingual; improved instruction-following and coding over Llama 3.1 70B.
Distributed under the Llama 3.3 Community License — same shape as the 3.1 / 3.2 licenses, >700M MAU carve-out preserved, naming convention requires “Llama” prefix on derivatives.
The last 3.x model. The next release was Llama 4 four months later.

Model

Llama 3.2 (Vision + Edge)

meta-llama/Llama-3.2-{1B, 3B, 11B-Vision, 90B-Vision}

Specialized

Available

Sep 25, 2024

First multimodal Llama (11B / 90B Vision) and first edge-targeted sizes (1B / 3B). EU regulators excluded from the multimodal license.

Released September 25, 2024 at Meta Connect 2024. Announcement: ai.meta.com.
Two product lines announced together: a vision-language pair (11B Vision built on Llama 3.1 8B; 90B Vision on Llama 3.1 70B) and an edge-targeted pair (1B and 3B text-only, distilled from larger 3.1 teachers, designed to run on phones and laptops).
128K-token context window across all four sizes.
EU multimodal carve-out: “the rights granted under the Llama 3.2 Community License Agreement are not being granted” to EU-domiciled individuals or EU-headquartered companies for the multimodal models, citing regulatory uncertainty under the EU AI Act, GDPR, and Digital Markets Act. EU users could still use the 1B / 3B text models. Slator coverage.
HuggingFace ids on the meta-llama org. Llama Guard 3 11B Vision shipped alongside as the multimodal-safety classifier.

Model

Llama Guard 3 family

meta-llama/Llama-Guard-3-{8B, 1B, 11B-Vision}

Specialized

Available

Jul 23, 2024

Three sizes shipped alongside Llama 3.1 / 3.2: 8B, 1B (edge), and 11B Vision (multimodal moderation). Input and output safety classification.

Llama Guard 3 8B shipped July 23, 2024 alongside Llama 3.1; the 1B and 11B Vision variants followed with Llama 3.2 on September 25, 2024.
Three sizes targeting different deployment surfaces — 8B for production servers, 1B for on-device, 11B Vision for multimodal-safety classification.
Input/output moderation: classifies prompts and responses against Meta's safety taxonomy (illegal violence, sexual content involving minors, etc.).
Distributed under the corresponding Llama 3.x Community License. Repo: github.com/meta-llama/PurpleLlama.

Model

Llama 3.1 (8B / 70B / 405B)

meta-llama/Llama-3.1-{8B, 70B, 405B}-Instruct

Flagship

Available

Jul 23, 2024

First “frontier-class” open-weights model. 405B competitive with GPT-4 / Claude 3.5 Sonnet at launch. 128K context. Co-released with Zuckerberg's open-source-AI letter.

Released July 23, 2024; the announcement is at ai.meta.com/blog/meta-llama-3-1.
First “frontier-class” open-weights model — Llama 3.1 405B was widely characterized as the first openly-available model competitive with GPT-4 / Claude 3.5 Sonnet on broad benchmarks at launch.
Three sizes (8B, 70B, 405B); 128K context across all three, up from 8K on Llama 3. Improved tool use and multilingual capability.
405B trained on ~16,000 H100 GPUs per Meta's accounting, ~3.8e25 FLOP total — the first Llama trained at this scale.
Co-released with Mark Zuckerberg's “Open Source AI Is the Path Forward” letter, which likened open-source AI's trajectory to Linux. The letter is now read with sharp dramatic irony given the closed-weights Muse Spark turn twenty-one months later (covered in the prose history below).

Model

Llama 3 (8B / 70B)

meta-llama/Meta-Llama-3-{8B, 70B}-Instruct

Flagship

Legacy

Apr 18, 2024

Two sizes. ~15T-token pretraining (7× Llama 2). 128K-vocab tokenizer. Grouped-query attention across all sizes. 8K context.

Released April 18, 2024; the announcement is at ai.meta.com/blog/meta-llama-3.
Two public sizes (8B and 70B); pretrained on ~15 trillion tokens — roughly seven times Llama 2's pretraining corpus, with four times more code data.
New 128K-vocabulary tokenizer; Grouped-Query Attention across all sizes (Llama 2 had GQA only at 70B).
8,192-token context window — bumped to 128K with Llama 3.1 three months later.
Distributed under the Meta Llama 3 Community License: same shape as Llama 2 (>700M MAU carve-out preserved); adds an attribution requirement that derivative model names include the “Llama 3” prefix.

Model

Code Llama 70B

codellama/CodeLlama-70b-{hf, Python-hf, Instruct-hf}

Specialized

Available

Jan 29, 2024

Largest Code Llama tier. Three flavors: base, Python specialization, Instruct. 16K context. The high-water mark of the Code Llama line.

Released January 29, 2024 as the largest Code Llama tier; followed the original 7B / 13B / 34B sizes by five months.
Three flavors per size: base, Python specialization, and Instruct (the chat-tuned coding assistant variant).
16,384-token context window — substantially longer than the 4K of the base Llama 2 line.
Distributed under the Llama 2 Community License. HuggingFace org: huggingface.co/codellama.
Marked the high-water mark of the named Code Llama line; later code-task work was folded into the general-purpose Llama 3.x and Llama 4 instruction-tuning.

Model

Purple Llama / Llama Guard 7B

meta-llama/LlamaGuard-7b

Specialized

Legacy

Dec 7, 2023

First Llama Guard. 7B classifier built on Llama 2. Launched alongside Purple Llama, Meta's umbrella for AI-safety tooling.

Released December 7, 2023 as the first Llama Guard; HuggingFace at meta-llama/LlamaGuard-7b.
7B classifier built on Llama 2 7B, fine-tuned for input and output safety classification (prompt and response moderation).
Launched as part of Purple Llama — Meta's umbrella for AI-safety tooling, also covering CyberSecEval (security benchmarks) and prompt-injection benchmarks. Repo: github.com/meta-llama/PurpleLlama.
Superseded by Llama Guard 3 family (mid-2024) and Llama Guard 4 (April 2025); status is Legacy on the “no longer the recommended classifier” reading.

Model

Code Llama (7B / 13B / 34B)

codellama/CodeLlama-{7b, 13b, 34b}-{hf, Python-hf, Instruct-hf}

Specialized

Legacy

Aug 24, 2023

Initial Code Llama release. Fine-tune of Llama 2 on code corpus. Three sizes × three flavors. 16K context.

Released August 24, 2023; the announcement is at ai.meta.com/blog/code-llama-large-language-model-coding.
Three sizes (7B, 13B, 34B), three flavors per size (base, Python, Instruct). The 70B tier followed five months later (January 2024 row above).
16,384-token context window — meaningfully longer than the 4K of the base Llama 2 line.
Distributed under the Llama 2 Community License (>700M MAU carve-out applies).
Superseded as the open-weights coding-model recommendation by general-purpose Llama 3.x / Llama 4 instruction-tuning by mid-2024.

Model

Llama 2 / Llama 2-Chat

meta-llama/Llama-2-{7b, 13b, 70b}, -chat-hf variants

Flagship

Legacy

Jul 18, 2023

First open-weights Llama with commercial use permitted. Three sizes (7B, 13B, 70B). 4K context. The Community License debuts with the >700M-MAU carve-out.

Released July 18, 2023 in partnership with Microsoft (Azure preferred partner); the announcement is at about.fb.com/news/2023/07/llama-2.
Three public sizes (7B, 13B, 70B); a 34B was trained but withheld for what Meta described as red-teaming reasons. Pretraining doubled to ~2T tokens vs. LLaMA 1.
4,096-token context window. The 70B tier introduced Grouped-Query Attention (GQA), the architectural choice that made later long-context expansions practical.
First Llama released as open weights with commercial use permitted, under the bespoke Llama 2 Community License Agreement. The license carves out companies whose products had >700 million monthly active users on the Llama 2 release date — a deliberate exclusion aimed at TikTok, Google, Apple, and other direct platform competitors. The carve-out has been preserved through every subsequent Llama license.
Llama 2-Chat ships RLHF-tuned chat variants in matching 7B / 13B / 70B sizes; the chat models powered most early third-party Llama deployments.
The release came four months after the LLaMA 1 leak and is widely read as Meta deciding the marginal cost of an official open-weights release was near zero once the weights were already in the wild — and that the upside (developer mindshare, ecosystem) was substantial.

The open-weights commercial era — July 18, 2023. Above this line: every Llama released as open weights with commercial use permitted, under successive Llama Community Licenses with the >700M-MAU carve-out preserved. Below: the original LLaMA 1, released as research-only by application — not actually open weights, and accidentally became so via the March 2023 leak that arguably forced the Llama 2 open release four months later.

Model

LLaMA 1 (7B / 13B / 33B / 65B)

research-only by application; weights leaked Mar 3, 2023

Flagship

Legacy

Feb 24, 2023

The original Llama. Four sizes. Research-only by application. Weights leaked to 4chan March 3, 2023, arguably forcing the Llama 2 open commercial release.

Released February 24, 2023; the paper is “LLaMA: Open and Efficient Foundation Language Models” (Touvron et al., 2023). The Meta page is at ai.meta.com/research/publications.
Four sizes (7B, 13B, 33B, 65B); pretraining 1–1.4 trillion tokens of fully public data. 2,048-token context window.
Demonstrated that state-of-the-art performance was achievable using only publicly-available data; LLaMA 65B competed with PaLM-540B and GPT-3 on broad benchmarks.
Released research-only, by application — non-commercial, gated by request form, available to academic and government applicants only. Not actually open weights.
The leak (March 3, 2023): a user on 4chan's /g/ board posted a torrent containing the LLaMA weights; the same day, a pull request was filed on Meta's official Llama GitHub repo proposing to add the magnet link to documentation. Meta filed DMCA takedowns through HuggingFace and GitHub through March 20. The leak is widely credited with forcing the Llama 2 open-commercial release four months later. The Register coverage; DeepLearning.AI's The Batch.

Click any row to expand. Each row has a stable id for sharing — e.g. /ai/llama/versions/#llama-4-maverick, #llama-3-1, #llama-2, #muse-spark. Llama family hub: llama.com; license texts at llama.com/license; HuggingFace org: huggingface.co/meta-llama.

The March 2023 LLaMA leak

Meta released LLaMA 1 on February 24, 2023 as a research-only model gated by application form, accessible to academic and government applicants and not under any open-source license. On March 3, 2023 — nine days later — a user posted a torrent containing the 7B and 65B weights to 4chan's /g/ technology board under the handle llamanon. The same day, a pull request was filed against Meta's official Llama GitHub repository proposing to add the magnet link to the README. The torrent reportedly pulled directly from Facebook's CDN at high speed, embedded with a unique download URL traceable to the leaker's gated-access grant.

Meta filed DMCA takedowns through HuggingFace and GitHub through March 20, 2023, but the weights had already propagated. Coverage at the time in The Register and DeepLearning.AI's The Batch documents the propagation timeline. Within weeks, the leaked LLaMA 1 weights were the de facto foundation for an enormous open-source fine-tune ecosystem — Alpaca, Vicuna, Guanaco, and dozens of others — on HuggingFace.

The leak is widely credited with forcing Meta's hand on Llama 2: with the weights already in the wild, the marginal cost of an official open-weights commercial release dropped to near zero, while the upside (developer mindshare, ecosystem effects, reduced lawsuit risk) was substantial. Llama 2 shipped with permissive commercial terms four months later. The decision-tree shape of the leak → commercial open release pattern is the load-bearing precondition for everything that followed in the Llama story through 2025.

The Llama Community License and the >700M-MAU carve-out

Llama 2 introduced the bespoke Llama 2 Community License Agreement — not an OSI-approved open-source license, but a Meta-authored license with permissive commercial terms and a single load-bearing carve-out: companies whose products had more than 700 million monthly active users on the Llama 2 release date (July 18, 2023) must request a separate license from Meta. The carve-out is a deliberate exclusion aimed at TikTok, Google, Apple, and other direct platform competitors; reading the precise language requires the actual PDF, which Meta hosts at llama.com/license and ai.meta.com/llama/license.

The carve-out has been preserved through every subsequent Llama license: Llama 3 Community License (April 2024) added an attribution requirement that derivative model names include the “Llama 3” prefix; the Llama 3.1 / 3.2 / 3.3 licenses extended the naming convention to require the “Llama” prefix on derivatives; the Llama 4 Community License preserved the same general shape. The Open Source Initiative has consistently said the Llama license does not meet the Open Source Definition because the use-restriction (any restriction tied to who you are, not what you do) is incompatible with the OSD — this is the substantive basis for the “source-available, not open source” framing of the line.

Llama 3.2 added a regional carve-out: the multimodal models (11B and 90B Vision) were excluded from EU-domiciled individuals and EU-headquartered companies, citing regulatory uncertainty under the EU AI Act, GDPR, and Digital Markets Act. The text-only 1B / 3B models remained available in the EU. Llama 4 carries the same EU exclusion, in identical wording — and because every Llama 4 model (Maverick, Scout, Behemoth) is natively multimodal, the restriction effectively covers the entire Llama 4 line for EU-domiciled individuals and EU-headquartered companies. EU-based end-users may still use Llama-4-powered services originating outside the EU: the restriction “does not apply to end users of a product or service that incorporates any such multimodal models.” A separate EU instrument arrived on July 28, 2026, when Meta said it would sign the EU AI Act Code of Practice on Transparency of AI-Generated Content — that one governs labeling and provenance of generated media (Content Seal is Meta's own implementation), not who may download and deploy the weights, so it leaves the license carve-outs untouched.

A sourcing detail worth knowing if you go looking for that language: the EU exclusion is not in the Community License text itself. It lives in the matching Acceptable Use Policy, which each license incorporates by reference, and it operates by withholding the Section 1(a) grant. The license PDF is silent on the EU; searching it alone will tell you the restriction doesn't exist. Re-verified 2026-10-19 against the Llama 4 Acceptable Use Policy and Community License as published in Meta's llama-models repo, alongside the Llama 3.2 equivalents — the Llama 4 wording is still identical to Llama 3.2's, end-user exemption included. The same pass confirmed the >700M-MAU carve-out intact in Llama 4, pegged to the Llama 4 version release date (April 5, 2025) and measured over the preceding calendar month.

The bespoke-license era has an end date, or at least an exception. Muse Glimmer, released August 10, 2026, ships under Apache 2.0 — a standard, OSI-approved license, and the first Meta frontier-lab model in the line's history not governed by a Meta-authored one. Practically, the two restrictions that have shaped every “can I ship with this?” question since July 2023 are simply absent: no >700M-MAU carve-out, so the platform competitors the Llama license was drawn to exclude are not excluded, and no EU carve-out, even though the model takes image input and would have tripped the multimodal restriction under a Llama license. The naming convention is gone too — derivatives need no “Llama” prefix. It is also published outside the Llama plumbing: the weights sit in a new meta-models HuggingFace org rather than meta-llama, and there is no Muse generation directory in the llama-models repo that hosts the per-generation license and use-policy texts. Whether Apache 2.0 would become the shape of future open Muse releases was the open question through September — Meta had announced on August 10, 2026 that an open-weights Muse Spark was coming, restated it on August 20 and September 2, and never named a license.

The answer, on October 15, 2026, was a third shape. Muse Spark Open ships under the Muse Model License — a Meta-authored license again, but not the Llama one. It grants use, modification, fine-tuning, distillation, distribution and sale, and attaches three conditions: the license and its Usage Policy travel with every copy; products must prominently display “Built with Muse Spark Open”; and anyone offering the model as a hosted service whose aggregate gross revenue with affiliates exceeds $20 million in any twelve-month period must notify Meta and sign a separate agreement. Set against the Llama Community License, the carve-out moved from who you are (700 million monthly users on a release date) to what you sell (inference as a service, above a revenue line), and the attribution requirement moved from the derivative's name to the product's interface. There is no EU exclusion — the only European text is the usual entity designation — and no “Llama” naming rule. The download is gated on both HuggingFace and Meta's own site. So the line now carries three license shapes at once: Apache 2.0 on the local-tier Glimmer, the Muse Model License on the frontier-scale Spark Open, and nothing at all on the API models, which as of October 19, 2026 remain proprietary and closed-weights — Muse Wonder (closed, with its Full variant sold only inside Muse Code), the three hosted Muse Spark versions, Muse Image, and Muse Voice Transcribe. The consumer Muse agent that launched on September 8 changes nothing here either — it is a product built on the closed models, with no weights of its own to license — and neither do Muse Realtime Voice and Muse Realtime Avatar, introduced inside that agent on September 23, which Meta has not put on the Model API, published weights for, or attached a license to at all.

FAIR, GenAI, and the org evolution

Meta's AI research home through 2022 was FAIR (Fundamental AI Research), founded in 2013 under Yann LeCun as Chief AI Scientist and led from 2023 by Joelle Pineau. FAIR's mandate was long-horizon basic research; its publication culture was open and academically-styled. LLaMA 1 was a FAIR project.

After ChatGPT's November 2022 launch caught Meta flat-footed, the company formed a separate Generative AI organization (GenAI) in February 2023, led by VP Ahmad Al-Dahle (who departed Meta in January 2026 to become CTO of Airbnb) and tasked with shipping product-grade generative AI on a faster cadence than FAIR's research timeline. Llama 2 and onward shifted progressively into GenAI under Manohar Paluri; FAIR was pushed toward longer-horizon work as the shipping pace of the GenAI org accelerated through 2024.

Through late 2024 and into 2025, FAIR's relevance to Meta's shipping AI products visibly waned. Internal reporting in April 2025 described FAIR as “dying a slow death.” Joelle Pineau announced her departure on April 1, 2025, last day May 30, 2025, after eight years leading FAIR; she was the most prominent internal advocate for open-source releases. Her exit set up the structural reorganization that came two months later.

The Scale AI acqui-hire and Meta Superintelligence Labs (June 2025)

On June 12–13, 2025, Meta announced a $14.3 billion investment for a 49% non-voting stake in Scale AI, valuing Scale at over $29 billion. The deal was structurally an acqui-hire: its purpose was to bring Scale CEO Alexandr Wang (then 28) to Meta as Chief AI Officer. Wang stepped down as Scale CEO but remained on Scale's board. Coverage in CNBC and Fortune.

On June 30, 2025, Mark Zuckerberg announced the formation of Meta Superintelligence Labs (MSL) with Wang as Chief AI Officer leading the new unit and former GitHub CEO Nat Friedman leading AI products. FAIR and GenAI were both consolidated under MSL; a new TBD Lab sub-group was created to work on next-generation LLMs. CNBC published the internal memo. The reorganization was the structural endpoint of Pineau's departure two months earlier and of the Llama 4 Behemoth delay reporting through May 2025.

Friction between Meta and Scale AI emerged through August 2025 as Scale's enterprise-data-labeling business saw revenue impact (competitors stopped sending training data to a Meta-owned vendor); TechCrunch covered the cracks in the partnership. MSL itself restructured into four sub-groups in August 2025.

Yann LeCun departs (November 2025) — AMI Labs

On November 19, 2025, Chief AI Scientist Yann LeCun announced his departure from Meta to launch his own startup focused on Advanced Machine Intelligence and “world models.” Reporting in CNBC and Bloomberg tied the decision to LeCun being asked to report to Wang — a structural demotion, since Wang's closed-source product orientation conflicts with LeCun's long-held public position that scaling LLMs is not a path to AGI and that open research is foundational to the field.

LeCun named the startup AMI Labs (Advanced Machine Intelligence) on December 18, 2025 and launched the public-facing site at amilabs.xyz in January 2026. On March 9, 2026 AMI announced a $1.03 billion seed round at a $3.5 billion pre-money valuation, co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions. The stated mission is “intelligent systems that understand the real world” rather than language-only LLMs. Meta is not an investor but a partnership is reportedly under discussion (LeCun has publicly suggested Meta could be AMI's first client), and AMI has a clinical-AI partnership with Nabla. With both Pineau and LeCun out, the open-research mandate inside Meta effectively ended at MSL's June 2025 formation.

The closed-weights turn — Muse Spark (April 2026)

On April 8, 2026, Meta Superintelligence Labs released Muse Spark, its first model since the June 2025 reorganization. Muse Spark shipped closed-weights, API-only, in private preview — the first frontier-AI model from Meta released without open weights since Llama 2 in July 2023. Branded under the new Muse family rather than Llama. The announcement is at ai.meta.com/blog/introducing-muse-spark-msl; coverage in CNBC and VentureBeat. Meta has stated it “hopes to open-source future versions” of Muse, but the launch model is proprietary.

Muse Spark is natively multimodal and ships with three operating modes — Instant (fast responses, no extended reasoning), Thinking (chain-of-thought), and Contemplating, which orchestrates parallel sub-agents (“thinking wider, not longer”), positioned competitively against Gemini Deep Think and the OpenAI o-series. Several outlets characterized the launch as the end of Meta's open-weights frontier-AI era. The closed-weights turn is the dramatic reversal of Mark Zuckerberg's July 2024 “Open Source AI Is the Path Forward” letter, which had argued (twenty-one months earlier) that open-source AI's trajectory would mirror Linux's path from inferior-but-cheap to dominant-via-ecosystem.

The line has since moved twice in one week without reversing course on weights. On July 7, 2026 MSL shipped Muse Image and previewed Muse Video (announcement), both closed-weights. On July 9, 2026 it shipped Muse Spark 1.1 (announcement) and launched a public preview of the Meta Model API. That API is the one partial opening so far: the original Muse Spark was private-preview only, and Muse Spark 1.1 is the first Meta frontier model any outside developer can call. It arrived three days after Meta switched off the equivalent service for its open models: the public preview of the Llama API, the first-party hosted endpoint for Llama 4 that Meta announced at LlamaCon in April 2025, was retired on July 6, 2026 (per Promptfoo's provider docs, which link Meta's deprecation notice; Meta has since taken that notice down). Access, not weights — there was still no HuggingFace release, and Meta's “hopes to open-source future versions” framing from April had not yet been acted on. The pattern repeated on July 24, 2026, when Meta wired the model's planning and tool use into Meta AI itself as consumer features — scheduled tasks, calendar and email connections, research syntheses, generated slides (newsroom post). Wider reach, same posture on weights.

On August 5, 2026 the line moved again, in a direction that says something about what Meta is selling. MSL shipped Muse Spark 1.2 — a coding-focused update — together with Muse Code, a terminal coding agent the model was co-trained with, shipped in beta (announcement). Weights stayed closed, but the commercial terms opened a new seam: alongside the standard API tier, Meta introduced a muse-spark-1.2-contributor tier at roughly a twelfth the input price and a twenty-first the output price, in exchange for permission to train future Meta models on the customer's prompts and completions. Where the Llama era traded weights for ecosystem, the Muse era is trading price for training data. The announcement also moved hosts: it was published on research.meta.ai, which now carries the Muse-line posts, and llama.com now redirects to developer.meta.com/ai/.

Five days later the turn partly reversed. On August 10, 2026 MSL released Muse Glimmer, a 30-billion-parameter local agent model, with open weights under Apache 2.0 (announcement) — the first downloadable model from Meta's frontier lab since Llama 4 in April 2025, and the first ever under a standard open-source license rather than a bespoke Llama Community License. Meta framed it as “keeping with our long tradition of sharing fundamental AI research.” The reopening is real but bounded: Glimmer is distilled from Muse Spark and sits a size class below it, and the flagship Muse Spark models stayed closed and API-only. What it does settle is the April 2026 “hopes to open-source future versions” framing — that has now been acted on once, at the small end of the line.

The same announcement went further than the model it shipped. On August 10, 2026, Chief AI Officer Alexandr Wang said Meta would also release open weights for a version of Muse Spark 1.2 — the flagship — “soon,” with no date attached. Meta restated the plan in its own words on August 20, 2026, opening a research post on the model's multimodal evaluations with the line that it was publishing them “ahead of the open-weights release” (announcement). That would move the reopening from the local tier to frontier scale, which is the question Glimmer left open. Through September it had not happened — the meta-models org held only the four Muse Glimmer repos, and Meta named neither a date nor a license. It happened on October 15, 2026, sixty-six days after the first promise, and is covered below.

The same day, Mark Zuckerberg published “The Future is for Everyone”, a letter proposing a philosophy of “individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety,” and asking whether superintelligence “will be centralized and restricted to a few institutions, or will it be a tool that empowers everyone.” It is the closest thing to a successor to the July 2024 open-source letter, and it landed alongside the first open-weights release since the turn — but it argues from access and distribution rather than from open weights, and it does not mention Llama, Muse, or open weights anywhere in its text.

What actually moved next was access again. On August 26, 2026, sixteen days after the open-weights promise and seven weeks after the model itself launched, Meta put Muse Image on the Meta Model API as muse-image-1.0 at $0.01 per image, resolving the “still evaluating whether it will make Muse Image available to outside developers” line from July. The weights stayed closed. That is now the line's settled rhythm at the frontier: each surface opens to developers as a metered endpoint, months after it opens to consumers, and never as a download. Worth noting where the announcement landed — a post on the AI developer blog, with nothing on research.meta.ai, ai.meta.com, or the newsroom. Meta now ships platform news and research news to different addresses.

Then, on August 31, 2026, the packaging changed rather than the model. Meta took Muse Code out of beta — four weeks after shipping it — and put flat-rate monthly subscriptions behind it, in three tiers (Everyday, High, and Power Usage) sold through Meta's Accounts Center and metered in requests per five hours instead of tokens (developer blog, subscription docs). The same release added staged multi-agent workflows, messaging between live sessions, and a conversation rewind. That gives the same closed model three distinct commercial shapes in under a month: metered tokens at the standard rate, tokens at roughly a twelfth the price if Meta may train on them, and a flat monthly fee for the agent that drives them. Through the end of August the shape of the arc was hard to miss: every change Meta had shipped at the frontier since the April 2026 launch was about how you buy access, and the one announced change to who may hold the weights, promised on August 10, had not landed. Where the announcement went said the same thing: the developer blog again, with nothing on research.meta.ai or the newsroom.

Then the models resumed. On September 1, 2026 MSL shipped Muse Voice Transcribe, its first real-time audio-perception model (announcement) — a new modality for the line, and its first product priced per minute of audio rather than per token. A day later, on September 2, 2026, it shipped Muse Spark 1.3 (announcement), an agentic-workflow update to the flagship. Both are closed-weights, so the posture is unchanged — but two things about the 1.3 launch are worth recording. It shipped incomplete, and was completed quietly two days later: Meta held its top “max reasoning” mode back “shortly after we finish additional safety testing,” so for two days the flagship was generally available while its highest setting was not — and, as VentureBeat pointed out on September 3, several of the launch's strongest disclosed scores belonged to that unshipped configuration rather than to the xhigh setting developers could actually call. On September 4, 2026 max went live on the Meta Model API and in Muse Code, on the standard tier only. Meta marked the change by editing the September 2 post in place rather than publishing a new one, and left both the post's published and its modified date reading September 2 — so the release is invisible to anyone diffing headlines, dates, or the blog index, and only the reasoning docs and the post body carry it. Nothing was concealed about the numbers; what was hard to see was the moment the gap closed. And it shipped narrower on one axis: Meta's own docs flag audio understanding on 1.3 as not fully supported and send audio work back to Muse Spark 1.2 or to the new transcription model. The open-weights promise was restated a third time in the same post — “bigger models, the Muse Spark open weights release, and more” — and this time without naming a version, so it is no longer clear that what was promised on August 10 for Muse Spark 1.2 is still what is coming. Zuckerberg restated it himself on X the same day, also without a version and in the plural — open-weights Muse Spark “releases” — which moved the promise up to the CEO without moving it any closer to a date (The Register). The one time frame anyone attached to it — “in the coming weeks,” per VentureBeat's account of the August 10 statement — came and went. Through the end of September, five months in, the Muse line had published open weights exactly once, at the size class furthest from the frontier.

Six days after that, the brand landed on a product. On September 8, 2026 Meta launched Muse — a consumer personal AI agent, carrying the family name itself, running on Muse Spark 1.3 (newsroom). Each user gets a dedicated Linux VM with a real browser; a separate host-side agent called Sentinel, which the model cannot override, approves every connector action and every network request, and the model is never shown a real credential (engineering writeup). It is free for most use with $20 and $100 monthly tiers above it, it carries no advertising, and Meta says a confidential-VM mode it cannot itself read is coming before the end of the year. For this page the disposition is the same as Muse Code's: a harness with no weights and no model id gets no row. But it is the clearest statement yet of what the closed-weights turn bought. Every Llama release was, in the end, a download. Everything the Muse line has shipped since April 2026 is a service you reach across Meta's network — a metered API, a data-for-discount tier, a subscription CLI, now a consumer agent running on a Meta-operated VM — and the one thing Meta has promised at the frontier and not delivered is still the thing the Llama era handed over by default. The name is now doing triple duty — a model family, a coding CLI, and a consumer app — which is worth keeping straight: on this page “Muse” alone means the agent, and the models carry their full names.

The following week the packaging moved twice more. On September 15, 2026 Meta launched Meta One, a consumer subscription spanning Instagram, Facebook, WhatsApp and Meta AI, whose Core ($7.99 a month) and Premium ($19.99 a month) bundles are sold partly on headroom for “creating and editing images and generating videos powered by Muse models” (newsroom) — the line reaching consumers as a metered allowance inside apps they already had, and a fifth distinct way to pay for models nobody can download. Three days later, on September 18, 2026, Meta opened the consumer agent's connector layer to outside developers at muse.ai/platform, and put SAM 3.1 on the Meta Model API as sam-3.1, at $2.50 per 1,000 images and $0.20 per 1,000 video frames (docs). SAM is a FAIR-line segmentation model rather than a Muse one, so it earns no row here — but it is the first non-Muse family that API has sold, and it runs the pattern backwards: Meta's own docs describe the SAM family as open-weight research models, and the SAM 3.1 checkpoint has been downloadable since March 27, 2026 (announcement), so here the metered endpoint arrives on top of the weights instead of in place of them.

Then the whole quarter's argument arrived on a stage. At Meta Connect on September 23, 2026, Meta Superintelligence Labs introduced two new models — Muse Realtime Voice, a real-time conversational speech model, and Muse Realtime Avatar, which consumes the same speech-token stream and generates synchronized video at roughly 870 ms from end-of-turn to first byte (research post, Connect recap). They are genuine frontier-lab models with a published architecture and published evaluations against Runway and HeyGen — and they are the first Muse models that arrived with no way to call them at all. Muse Spark went closed-weights but got an API; Muse Image and Muse Voice Transcribe followed it onto that API; Muse Glimmer went the other way and became a download. These two are neither: Meta's own Model API roster still lists five families and names neither of them, there is no model id, no endpoint, no price, and no HuggingFace release. The only way to reach them is to talk to the consumer app. The same keynote pushed that app onto more surfaces rather than fewer — a Mac build with computer use, AI glasses “in the coming months,” an email address of its own, a pocket voice device called Muse Charm with “more to share later this year,” and a much longer connector roster including Walmart, Best Buy, Instacart, Expedia, GitHub and Notion. Meanwhile the one thing promised at the frontier and not delivered went unmentioned: Connect restated the open-weights Muse Spark release nowhere, and the meta-models org still held only the four Muse Glimmer repos the next morning.

The week after Connect the selling moved up a level, from products to an organization. On September 28, 2026 Zuckerberg announced Meta Enterprise Platform, which he called “the next major pillar of our business,” to take Meta's AI stack — the Muse agent, Meta Business Agent, what his post calls the Muse API, and Muse Code — to businesses and developers, and hired former MongoDB CEO Chirantan “CJ” Desai to run it as Chief Enterprise Platform Officer, reporting directly to him (newsroom). The next day Meta launched Muse for Small Business, a set of business skills and connectors inside the consumer agent, which by then was available in Canada as well as the US (newsroom). Neither announcement mentioned weights. Every item the enterprise post names is something you reach across Meta's network; neither Muse Glimmer nor Llama is named.

October opened with the rulebook rather than a model. On October 2, 2026 Meta published an update to what it now calls the Meta Superintelligence Scaling Framework, the policy it says governs how its most capable models are trained, evaluated and released (research post). The update extends its loss-of-control requirements to containment during training and evaluation, expands the risk considerations it applies to open-weight releases specifically, and announces a new AI committee of Meta's board to review future changes; Meta tied it to commitments made at the White House that week. Two releases followed inside two weeks, and each cites the framework.

On October 12, 2026 MSL shipped Muse Wonder and Muse Wonder Full, “our most capable models to date, built on the groundwork we laid with Muse Spark” (newsroom, research post) — the first model to move the frontier above Muse Spark since April. Closed-weights, and sold in a shape the line had not used before: Muse Wonder on the Meta Model API at $4 / $20 per million tokens with no contributor tier, and Muse Wonder Full — the better of the two — nowhere on the API at all, available only through Muse Code's two top subscription plans. Where the Llama era's best model was a download and the early Muse era's best model was a metered endpoint, the best model Meta now sells is bundled with its coding agent. Meta's docs still call Muse Wonder a preview.

Three days later, on October 15, 2026, the promise from August landed. MSL released Muse Spark Open — 816 billion parameters, 42 billion active, the same base as Muse Spark 1.3 — as downloadable weights on HuggingFace (announcement), calling it “the most capable US open-weight model to date.” It is Meta's first frontier-scale open-weights release since Llama 3.1 405B in July 2024, and the first since the closed-weights turn. The license is neither Apache 2.0 nor a Llama Community License but a new Muse Model License — “limited commercial terms, which allow many developers to use the model for free,” in Meta's words, with a $20 million-revenue clause for hosted-inference sellers and a required “Built with Muse Spark Open” credit (details in the license section above). Three things about the timing are worth recording. The promise was made for Muse Spark 1.2, restated without a version, and fulfilled with 1.3's base — the version moved on while the words stayed the same. The repos existed on HuggingFace from September 15 as private, gated repositories, so a month of “nothing published” was true of what the public could see and not of what Meta had staged. And the weights arrived three days after Muse Wonder superseded the model they reproduce, which gives the Muse line a posture the Llama line never had: the newest model closed and partly subscription-only, the previous generation open under a Meta-authored license with conditions, and the local tier open under Apache 2.0. The open weights trail the closed frontier by one generation.

The Muse line is treated on this page as Llama's successor — same Meta organization (MSL), same flagship-frontier slot — even though it is not Llama-branded. The open-versus-closed question that defined the Muse line's first six months now has a settled shape rather than an open one: Muse Glimmer answered it for the local tier in August, and Muse Spark Open answered it at frontier scale in October, each under a different license, while the newest model stayed closed. Whether that one-generation lag holds — whether a Muse Wonder Open follows, and under which terms — is the question going forward. No new Llama has shipped since April 2025 and none is announced; the reopening arrived under the Muse brand and two different licenses instead.

Where to run Llama

Llama is the most widely-distributed frontier-AI line because the weights are open. Inference paths through 2025–2026 break into three categories.

Self-host. Download from the HuggingFace meta-llama org and run with vLLM, llama.cpp, Ollama, MLX (Apple Silicon), or TensorRT-LLM. The 1B / 3B Llama 3.2 edge models are specifically designed for on-device deployment. Since August 10, 2026 the self-host path also covers a Muse model: Muse Glimmer 30B ships open weights under Apache 2.0 from the separate meta-models org, quantized to fit a single 24 GB or 32 GB consumer GPU, with LM Studio, Unsloth, ExecuTorch and SGLang added to the runtime list. Since October 15, 2026 it also covers a frontier-scale one: Muse Spark Open, 816B parameters under the Muse Model License, from the same meta-models org (gated download) — about 537 GB of weights in NVFP4 or 1.65 TB in BF16, so “self-host” here means a four- or eight-GPU Blackwell node or a pair of M5 Ultra Mac Studios under MLX, not a desktop GPU.

Hosted-inference providers. Together AI, Fireworks AI, Groq (long notable for very high tokens-per-second on Llama 3 70B, but no longer a Llama host for most customers: its deprecation log shows Llama 4 Maverick shut down on March 9, 2026, Llama 4 Scout on July 17, and Llama 3.1 8B and Llama 3.3 70B on August 16, all in favor of OpenAI's and Alibaba's open models, with only committed-spend enterprise contracts exempt — though it is back for the open Muse, as a named launch partner for Muse Spark Open), Replicate, Perplexity Labs, Cerebras, SambaNova. Not every one of them still carries the newest weights: for Llama 4 Maverick, Artificial Analysis tracks six API hosts as of October 19, 2026 — Amazon Bedrock, Azure, DeepInfra, DigitalOcean, Novita and Parasail; Together AI has dropped off that list since late September. For Muse Spark Open the launch-partner hosts are Together AI, Fireworks AI, Baseten, Groq, Modal, Nebius and OpenRouter, plus the Amazon SageMaker, Google Cloud Model Garden and Microsoft Foundry catalogs. Pricing is typically a fraction of comparable closed-weights frontier-model API rates because providers compete on inference cost, not model rights.

Hyperscalers. AWS Bedrock, Azure AI (Microsoft was Meta's launch partner for Llama 2 in July 2023), Google Cloud Vertex AI, Oracle OCI, IBM watsonx. Meta's own Llama API (announced at LlamaCon 2025-04-29) was the first-party hosted option for the open-weights line until Meta retired its public preview on July 6, 2026 (per Promptfoo's provider docs, which link Meta's since-deleted deprecation notice); its developer portal now forwards to the Meta Model API docs, whose model roster lists no Llama model as of October 19, 2026. For hosted Llama, the third-party hosts named in this section are what remain. The Muse line now splits three ways. Its newest model is API-and-subscription only: the Meta Model API, which has served Muse models to outside developers since July 9, 2026, carries Muse Wonder (muse-wonder-1.0, since October 12, 2026, $4 / $20 per million tokens, standard tier only) above Muse Spark 1.3, 1.2 and 1.1, with the max reasoning setting live on Wonder and on 1.3's standard tier; Muse Wonder Full is reachable only through the Muse Code terminal agent on its Power or Maximum subscription plan. Muse Code is the first-party CLI on top of the API, out of beta since August 31, 2026, native on Windows as well as macOS and Linux since September 16, and billable either per token or on one of four flat monthly plans; OpenRouter is the first third-party route to the hosted models. The previous generation is a download: Muse Spark Open, since October 15, 2026, belongs to the self-host paragraph above and to the hosted-provider list, and is not on the Meta Model API at all. Muse Image joined the same API on August 26, 2026 as muse-image-1.0, at $0.01 per image, through /v1/images/generations and /v1/images/edits; OpenRouter routes it as meta/muse-image. Muse Voice Transcribe joined on September 1, 2026 as muse-voice-transcribe-1.0, on its own endpoints — a WebSocket at wss://api.meta.ai/v1/asr/realtime for live audio and POST /v1/asr/transcribe for recordings — billed at $0.18 per hour of audio rather than per token, and also wired into Meta AI for Mac and Muse Code for dictation. That API stopped being Muse-only on September 18, 2026, when Meta added SAM 3.1, its segmentation model, as sam-3.1 over POST /v1/responses — billed per image and per video frame rather than per token, and the one model on the API whose weights you can also just download. It is a FAIR-line model rather than a Muse one, so it has no row in the table above. Muse Glimmer, likewise, is not on that API — since August 10, 2026 it is a download, and belongs to the self-host paragraph above. Muse Image and the Muse Spark models are also live in the Meta AI app and on meta.ai, and since September 8, 2026 Muse Spark 1.3 also drives Muse, the consumer personal agent (Meta says Muse Wonder will power it next, subscribers first) — in the US and, since September 18, 2026, Canada, on iOS, Android, muse.ai, and inside WhatsApp, each user's agent running on its own Meta-hosted Linux VM, plus a Mac app since September 17, 2026 that can drive native applications with permission, and Meta's AI glasses promised “in the coming months” as of Connect on September 23. That is a way to use the model, not a way to run it: there is no endpoint, no model id, and no download behind it. The same is true twice over of Muse Realtime Voice and Muse Realtime Avatar, the conversational-speech and embodiment models Meta introduced on September 23 — they are not on the Meta Model API, not on HuggingFace, and not routed by anyone; the consumer app is the only place they exist. Note that llama.com now redirects to developer.meta.com/ai/, which as of October 19, 2026 forwards on again to dev.meta.ai, the Meta Model API portal — Meta's model pages and developer blog have moved to that host too — and the license and use-policy paths redirect with it.

People who shaped Llama

Mark Zuckerberg — CEO of Meta. The 2024 “Open Source AI Is the Path Forward” letter and the 2025 strategic decisions (Scale AI acqui-hire, MSL formation, Muse Spark launch) all run through Zuckerberg's office. The closed-weights turn reverses the public position the letter staked out twenty-one months earlier.

Yann LeCun — Chief AI Scientist 2013–2025, FAIR cofounder. 2018 Turing Award. Departed November 2025 to launch AMI Labs (Advanced Machine Intelligence), a world-models startup that emerged from stealth in January 2026 and closed a $1.03B seed at a $3.5B pre-money valuation on March 9, 2026.

Joelle Pineau — led FAIR 2023–2025. The internal advocate for open-source releases through the Llama 1–3 era. Departed May 2025 ahead of the MSL reorganization.

Ahmad Al-Dahle — VP of Generative AI at Meta through Llama 2 / 3 / 4; spokesman for the LMArena Maverick episode in April 2025. Departed Meta in January 2026 to become CTO of Airbnb (Airbnb announcement, January 14, 2026).

Alexandr Wang — Chief AI Officer of Meta since June 2025; head of Meta Superintelligence Labs. Previously CEO of Scale AI. The architect of the closed-weights pivot and the Muse line.

Nat Friedman — head of AI products at MSL since June 2025; previously CEO of GitHub.

Shengjia Zhao — Chief Scientist of Meta Superintelligence Labs since July 25, 2025, setting the lab's research agenda alongside Zuckerberg and Wang. Previously at OpenAI, where he was a co-creator of ChatGPT and GPT-4 (CNBC, TechCrunch). His appointment four months before LeCun's departure left MSL's scientific direction with a hire from a closed-weights lab rather than with FAIR's founding generation.

The competitive landscape

Llama is — through April 2025 — the dominant open-weights frontier-AI line by deployment volume. The closest open-weights competitors are Mistral (French; mixed Apache 2.0 / Mistral Research License / proprietary tiers across the line, with the December 2025 “Mistral 3” family relaunch returning the open releases to Apache 2.0 — see Mistral Versions), DeepSeek (Chinese, MIT-licensed for the V3 / R1 line and onward, the December 2024 / January 2025 inflection — see DeepSeek Versions), Alibaba's Qwen (also fully open-weights and dominant on HuggingFace leaderboards through 2025 — see Qwen Versions), and xAI's Grok 1 (open-weights only for the first generation, see Grok Versions). The closed-weights frontier competitors — ChatGPT, Claude, Gemini — have all stayed closed-weights since their inception. With Muse Spark in April 2026, Meta is the only major frontier lab to have started open-weights and pivoted closed — and, with Muse Glimmer in August 2026 and Muse Spark Open in October 2026, the only one to have pivoted partway back, releasing a local-tier model under Apache 2.0 and then a frontier-scale one under its own Muse Model License while keeping its newest model, Muse Wonder, closed. Meta's own framing for Muse Spark Open is “the most capable US open-weight model to date,” a claim scoped by nationality against the Chinese open-weights lines above; this page does not attempt a benchmark roundup or a ranking. The open question for the line going forward is whether the one-generation lag between Meta's closed frontier and its open weights holds, and under which license the next open release ships.

Use Llama

The browser cannot detect which Llama model you've used or are using — there's no fingerprint or header that exposes it. The block below carries the practical information instead: the current open-weights model identifiers, a copy-paste self-host command, and the surfaces where Llama is available.

Current open-weights model identifiers

HuggingFace ids on the meta-llama org for the Llama line, and on the separate meta-models org for the open-weights Muse release. Verify against huggingface.co/meta-llama and huggingface.co/meta-models for the freshest list.

# Llama 4 — current open-weights flagship line
meta-llama/Llama-4-Maverick-17B-128E-Instruct
meta-llama/Llama-4-Scout-17B-16E-Instruct

# Llama 3.x — still widely served
meta-llama/Llama-3.3-70B-Instruct
meta-llama/Llama-3.1-{8B, 70B, 405B}-Instruct
meta-llama/Llama-3.2-{1B, 3B, 11B-Vision, 90B-Vision}-Instruct

# Specialized — coding and safety
meta-llama/Llama-Guard-4-12B
codellama/CodeLlama-70b-Instruct-hf

# Successor — open weights (separate HuggingFace org)
meta-models/Muse-Spark-Open-816B-A42B
meta-models/Muse-Spark-Open-816B-A42B-NVFP4
                 # 816B MoE, 42B active; 1M context; Muse Spark 1.3's base.
                 #   Muse Model License: no MAU/EU carve-out, but "Built with
                 #   Muse Spark Open" credit + $20M-revenue clause for MaaS.
                 #   Gated download. ~537 GB (NVFP4) / ~1.65 TB (BF16).
meta-models/Muse-Glimmer-30B
                 # 30B local agent model; 131K context; Apache 2.0; also -GGUF,
                 #   -assistant, -ExecuTorch-PTE. No MAU or EU carve-out.

# Successor — closed-weights, API only (no HuggingFace release)
muse-wonder-1.0  # current frontier (preview); Meta Model API; $4 / $20 per M;
                 #   no contributor tier; reasoning_effort up to "max".
                 #   Muse Wonder Full: no API id — Muse Code Power/Maximum plans only
muse-spark-1.3   # cost tier under Wonder; "recommended for new work" Spark.
                 #   base URL https://api.meta.ai/v1 — docs at dev.meta.ai/docs
                 #   reasoning_effort goes up to "max" (live 2026-09-04);
                 #   standard tier only — not on the -contributor tier
muse-spark-1.3-contributor
                 # cheaper tier — Meta trains on your prompts + completions
muse-spark-1.2   # superseded by 1.3, but still the one to use for AUDIO in;
                 #   1.3's audio support is flagged "not fully supported"
muse-spark-1.2-contributor
muse-spark-1.1   # superseded; still served
muse-voice-transcribe-1.0
                 # streaming speech-to-text + diarization; since 2026-09-01
                 #   wss://api.meta.ai/v1/asr/realtime, POST /v1/asr/transcribe
                 #   $0.18 per hour of audio, not per token
muse-image-1.0   # text-to-image + editing; on the API since 2026-08-26
                 #   $0.01/image — /v1/images/generations, /v1/images/edits
Muse Spark       # superseded; research.meta.ai/blog/introducing-muse-spark-msl

Quick self-host (Ollama)

Ollama wraps the HuggingFace download and the inference loop. For higher throughput, vLLM and TensorRT-LLM are the routine production choices.

$ brew install ollama          # macOS; apt / curl install on Linux
$ ollama pull llama3.3:70b
$ ollama run  llama3.3:70b "Hello, Llama."

Where to run Llama

Three categories — self-host, hosted-inference providers, and hyperscalers. Pricing varies by orders of magnitude; the open weights are the same across all of them.

# Self-host runtimes
https://ollama.com/                         # single-binary, easiest entry
https://github.com/ggerganov/llama.cpp      # CPU + GPU, edge-friendly
https://github.com/vllm-project/vllm        # production-grade throughput

# Hosted-inference providers
https://www.together.ai/
https://fireworks.ai/
https://groq.com/                            # dropped Llama for most tiers by 2026-08-16; hosts Muse Spark Open
https://replicate.com/

# Hyperscalers
AWS Bedrock, Azure AI, Google Cloud Vertex AI, Oracle OCI, IBM watsonx

# Meta's first-party Llama API — retired 2026-07-06
https://www.llama.com/                       # now forwards to dev.meta.ai (Muse / Model API)

Licensing

Each Llama generation ships with a successor “Llama Community License.” Read the actual PDF before shipping at scale — the >700M-MAU carve-out persists, and Llama 3.2 multimodal carries an EU restriction. Muse Glimmer escapes both as plain Apache 2.0; Muse Spark Open escapes both too, but trades them for a revenue clause and an attribution line under Meta's own Muse Model License.

# Authoritative license texts
https://www.llama.com/license/
https://ai.meta.com/llama/license/

# Acceptable Use Policy applies to every Llama license
https://www.llama.com/use-policy/

# Notable carve-outs
>700M MAU on the release date → separate Meta license required
Llama 3.2 multimodal models → not granted to EU-domiciled users
Llama 4 (all models, natively multimodal) → same EU exclusion applies
Muse Spark     → Muse Model License (2026-10-15) — open weights, NOT
  Open           Apache 2.0: use/modify/distill/distribute/sell allowed;
                 must display “Built with Muse Spark Open”; selling it
                 as a service above $20M aggregate revenue in any 12
                 months → notify Meta + separate agreement. No MAU
                 carve-out, no EU exclusion, no naming rule. Gated download.
Muse Glimmer   → Apache 2.0 open weights — NO MAU carve-out,
                 NO EU carve-out, no “Llama” naming requirement
Muse Wonder    → proprietary, closed-weights; Meta Model API, standard
                 tier only (no contributor tier). Muse Wonder Full →
                 not on the API at all; Muse Code Power/Maximum plans only
Muse Spark 1.3 → proprietary, closed-weights; Meta Model API
                 standard tier → Meta does not train on your prompts
                 -contributor tier → cheaper; Meta DOES train on them
                 its base shipped as open weights 2026-10-15 (Muse Spark
                 Open, above) — the hosted model itself stays closed
Muse Spark 1.2 → proprietary, closed-weights; superseded by 1.3
Muse Spark 1.1 → proprietary, closed-weights; superseded
Muse Voice     → proprietary, closed-weights; Meta Model API,
  Transcribe     metered per hour of audio. No contributor tier.
Muse Spark     → proprietary, closed-weights, no redistribution
Muse Image     → proprietary, closed-weights; in-app and, since
                 2026-08-26, metered on the Meta Model API
Muse Code      → agent, not a model — no weights to license.
                 Out of beta 2026-08-31; pay per token, or one of four
                 flat monthly plans bound to the CLI's own key (plans
                 default to Contributor mode — Meta trains on your
                 content unless you switch it off)
Muse           → consumer personal agent (2026-09-08), not a model —
  (the app)      no weights to license. Runs Muse Spark 1.3 on a
                 per-user Meta-hosted VM. Free tier, $20/mo, $100/mo.
                 Mac app since 2026-09-17; AI glasses "coming months"
Muse Realtime  → introduced 2026-09-23 inside the Muse app. No model
  Voice /        id, no endpoint, no price, no HuggingFace release —
  Avatar         not on the Meta Model API roster at all, so there is
                 no license surface to read. Not rows in the table.

Sources: Llama family hub; per-release announcement posts at research.meta.ai/blog, ai.meta.com/blog, developer.meta.com/ai/resources/blog, and about.fb.com/news; Meta Model API docs at dev.meta.ai/docs; HuggingFace orgs meta-llama, codellama, and meta-models; license texts at llama.com/license; Purple Llama repo at github.com/meta-llama/PurpleLlama; contemporaneous reporting in NYT, WSJ, Bloomberg, The Information, CNBC, NPR, TechCrunch, The Register, Fortune, and DeepLearning.AI. Last updated October 19, 2026.

Mungomash LLC · More AI pages

Llama Versions

Family & status

Llama version table

The March 2023 LLaMA leak

The Llama Community License and the >700M-MAU carve-out

FAIR, GenAI, and the org evolution

The Scale AI acqui-hire and Meta Superintelligence Labs (June 2025)

Yann LeCun departs (November 2025) — AMI Labs

The closed-weights turn — Muse Spark (April 2026)

Where to run Llama

People who shaped Llama

The competitive landscape

Use Llama

Current open-weights model identifiers

Quick self-host (Ollama)

Where to run Llama

Licensing