2023 – 2026

Qwen Versions

The latest generally available Qwen flagship is Qwen3.8-Max (August 3, 2026) — a sparse-MoE model with 2.4 trillion parameters, 95 billion active, native vision-language input, a 1,000,000-token context window, and hybrid thinking on by default. It sits first on Qwen Cloud's capability-ordered model list, ahead of the mid-tier Qwen3.7-Plus (May 31, 2026) and the cost tier Qwen3.8-Flash (August 26, 2026); the prior Max-tier flagship, Qwen3.7-Max (May 20, 2026), is still served but no longer leads. The promise Alibaba made at the August 3 launch has now landed: on August 12, 2026 it published Qwen/Qwen3.8-2.4T-A95B, the first open-weights release of a Qwen-Max-class model and the largest open-weights language model published to date — though under a bespoke Qwen3.8-Max License with revenue-share conditions rather than Apache 2.0, and text-only (the hosted qwen3.8-max keeps the vision input). Two days later, on August 14, 2026, the companion Qwen3.8-27B dense checkpoint shipped under plain Apache 2.0. The newest open-weights model in the flagship line is Qwen3.8-Flash-Next (August 26, 2026), a 125B model activating 6B parameters that Alibaba published as an explicit early preview of the architecture Qwen4 will be built on — and which carries a fifth license, the Qwen Community License 1.0. The newest open-weights release of any kind is Qwen-Drive-1.0-4B (August 28, 2026), the line's first autonomous-driving model, under plain Apache 2.0. I track every Qwen / Tongyi Qianwen release here — from Qwen-7B in August 2023 onward — with HuggingFace ids, ship dates, family (Flagship / Reasoning / Specialized), and license terms (Apache 2.0 / Tongyi-Qianwen / the new Qwen3.8-Max License / proprietary). Below the table: the April 2023 Tongyi launch as Alibaba's ChatGPT response, the licensing turn at Qwen2, the QwQ reasoning track, the Qwen3 hybrid-reasoning era, the U.S. chip export-control context, and the HuggingFace-leaderboard dominance through 2025–2026.

Family & status

Family

Flagship — the main Qwen-N chat lineage including the proprietary Qwen-Max tier
Reasoning — the QwQ line; converged into the Flagship V-series at Qwen3's hybrid Thinking / Non-Thinking architecture
Specialized — Qwen-Coder, Qwen-VL, Qwen-Audio, Qwen-Math, Qwen-Omni, Qwen-Image

Status

Current — actively recommended; the latest in its product slot
Available — weights still served via HuggingFace, or proprietary API still served on DashScope, but superseded
Legacy — deprecated, retired from DashScope, or no longer recommended

Qwen version table

Model
Qwen-Drive-1.0
Qwen/Qwen-Drive-1.0-4B — open weights (Apache 2.0); no hosted model string
Specialized
Current
Aug 28, 2026
Qwen's first model built for autonomous driving, and the first release in the line's physical-AI push that actually ships weights — the June 2026 Qwen-Robot Suite is still an undisclosed enterprise pilot. The Qwen3.5-4B vision-language model left architecturally untouched, with two modules attached: a BEV perception head doing 3D detection, occupancy prediction and map segmentation, and a Planning Expert that generates ego trajectories by flow matching. Apache 2.0, on HuggingFace and ModelScope.
  • Weights published August 28, 2026 — the Upload folder using huggingface_hub commit on Qwen/Qwen-Drive-1.0-4B, matching the single upload commit on the ModelScope mirror (both land on August 28 in Alibaba's own time zone). Code and docs are at QwenLM/Qwen-Drive-1.0, public since August 25; the technical report is arXiv 2609.00111, submitted August 31.
  • A staged release, and the staging is visible. The repositories were populated on August 28 but neither resolved publicly for several days — when I probed them on September 2 both HuggingFace and ModelScope still answered as though the model did not exist, and that is the same day the model card took its last two commits; the card's first community thread is dated September 3. So the upload date and the date you could first download it are not the same date here. This one also never got a per-release announcement: the README links the generic qwen.ai/research index rather than a post, and there is no Alibaba Cloud Community write-up.
  • The pretrained VLM is deliberately left alone. Alibaba's stated design goal is a driving model that adds capability without touching the base architecture: Qwen3.5-4B serves as the shared vision-language model, a BEV perception head is attached as an “explicit, inspectable” 3D probe (joint 3D object detection, semantic occupancy prediction, BEV map segmentation), and a Planning Expert conditions on the same shared representations to emit future ego trajectories through flow matching. The unchanged LLM decoder still answers free-form questions, so driving VQA and general VQA run on one model. The README claims it as the first vision-language foundation model for driving to unify 3D perception and VQA at the pretraining stage and then extend to planning — that framing is Alibaba's.
  • One directory, four pieces. The 9.1 GB VLM at the root (which on its own serves the VQA mode), plus planner-sft/ (2.1 GB, imitation-trained), planner-rl/ (2.1 GB, the same expert after reward optimization) and perception/ (0.5 GB). A head is chosen at load time via QwenDriveForPlanning.from_pretrained(…, planner=…); the config declares model_type: qwen_drive. Alibaba notes planner-rl was reward-optimized only on reasoning-conditioned rollouts, so it belongs in the reasoning planning mode, while planner-sft covers both.
  • Vendor benchmark numbers, reproduced under one protocol. Planning: 90.7 PDMS on NAVSIM v1.1 navtest for the RL planner (91.4 best-of-6) against 88.2 for the SFT planner, 7.91 RFS on the Waymo Open Dataset end-to-end test, and 0.34 m minADE at 3 s on NVIDIA's PhysicalAI open-loop set. Driving VQA: 77.8 on LingoQA, 7.78 Ego3D RMSE, 66.1 SURDS, 74.5 WaymoQA overall — all well above the Qwen3.5-4B base and the InternVL / Cosmos / MiMo-Embodied comparators in Alibaba's table. The more interesting number is what did not move: general VQA stays roughly at the base model (75.9 vs. 75.3 MMStar, 72.7 vs. 73.4 MMMU), which is the whole point of not touching the architecture. These are Alibaba's own evaluations.
  • License: Apache 2.0 — stated in the model card's YAML frontmatter (license: apache-2.0) and in the GitHub README's License section, so this one sits in the permissive tier rather than with the bespoke August licenses on Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next. Open-weights only: there is no hosted counterpart — qwencloud.com/models/qwen-drive-1.0 returns 404, and no qwen-drive string appears in the Model Studio billing table (both checked September 4, 2026).
Model
Qwen3.8-Flash-Next
Qwen/Qwen3.8-Flash-Next — open weights · qwen3.8-flash — hosted, Qwen Cloud / Model Studio
Flagship
Current
Aug 26, 2026
An explicit early preview of the architecture Qwen4 will be built on, shipped as working open weights rather than a paper. 125B total with only 6B active, plus a 51B n-gram embedding table; native vision-language, 262K context extensible to 1M. Four architectural swaps at once — Qwen Sparse Attention, Gated Residual, n-gram embedding, a Muon-based training recipe. Not Apache 2.0: it introduces a fifth license tier, the Qwen Community License 1.0. Hosted the same day as qwen3.8-flash at $0.15 / $0.47 per million.
  • Released August 26, 2026 — the date in the News log of QwenLM/Qwen3.8-Flash-Next, matching the Upload folder commit on the HuggingFace card Qwen/Qwen3.8-Flash-Next (an FP8 build landed alongside it, and ModelScope carries both). Announcement at qwen.ai/blog with a technical report; coverage in MarkTechPost, TechNode, and The Decoder. Alibaba's own English write-up, Alibaba Releases Qwen3.8-Flash (August 27, 2026), files it under the hosted name and confirms the 125B / 6B-active / 51B-n-gram split, the 262K-to-1M context, and the 1 RMB / 3 RMB per-million pricing.
  • A Qwen4 architecture preview, said out loud. The config declares Qwen4ExpForConditionalGeneration / qwen4_exp, and Alibaba's README states the intent plainly: it “plays the same role that Qwen3-Next played for Qwen3.5” — release the architectural changes early so the community can examine them before the full model family is built on top. The Gated DeltaNet + Gated Attention design that Qwen3-Next previewed in September 2025 did go on to carry Qwen3.5, 3.6, 3.7, and 3.8, so the precedent is a real one rather than a marketing frame.
  • Four changes at once — attention, residual, embedding, optimizer. Qwen Sparse Attention (QSA) replaces Gated Attention in the hybrid: a lightweight indexer picks relevant context at micro-block rather than per-token granularity, on a budget of 512 blocks / 2,048 tokens. Gated Residual widens the residual stream to four branches (bottleneck rank 320) with data-dependent read and per-branch write gates. N-gram Embedding adds a 20-million-entry bigram/trigram table at layer 2, scaling capacity for almost no extra compute and offloadable to host memory with async prefetch. And the training recipe splits Muon and AdamW across weight categories, dropping batch-size warmup entirely against a refitted scaling law.
  • 125B total, 6B activated, plus 51B of n-gram embeddings and 4B of MTP. 48 layers, hidden dimension 2,560, 512 experts with 10 routed + 1 shared active, laid out as 12 × (3 × Gated DeltaNet → MoE, then 1 × QSA → MoE). Unlike the text-only Qwen3.8-2.4T-A95B, this one keeps the vision encoder — it is a causal LM with vision input, 262,144-token context extensible to 1,000,000. Alibaba says training cost roughly one-ninth of Qwen3.7-Plus while scoring above it on coding and office work: 58.7 vs. 16.5 on DeepSWE 1.1 and 62.5 vs. 55.8 on SWE-bench Pro, per the model card's own table, which also puts it ahead of Qwen3.8-27B (42.2 / 61.7) and DeepSeek-V4-Flash-0731 (54.4 / 56.0). These are vendor numbers.
  • A fifth license tier: the Qwen Community License 1.0. The card reports license: other / license_name: qwen-community-1.0, and the repo's LICENSE is a near-copy of the Qwen3.8-Max License — MIT-shaped rights, the same attribution condition above 100 million MAU or $20 million monthly revenue, the same carve-outs defining “Model as a Service” and “AI Work Assistant” — with one substantive difference, and it cuts the stricter way: the Max license only requires a separate commercial license once a MaaS or AI-work-assistant licensee clears $50 million of trailing-twelve-month group revenue, while this one drops the revenue threshold entirely and applies to any such business at any size. Internal use is still carved out.
  • Hosted the same day as qwen3.8-flash. Per the model card, “Qwen3.8-Flash is the official version based on Qwen3.8-Flash-Next with more production features, e.g., 1M context length by default, official built-in tools.” The marketplace page lists $0.15 / $0.47 per million input / output tokens (implicit cache reads $0.016), 1M context, 991K max input, 131K max output, 262K max reasoning, 2M TPM and 15K RPM, with code_interpreter, web_search, web_extractor and the image-search tools built in — read August 28, 2026. On Qwen Cloud's capability-ordered Choose models page it now holds the cost tier, displacing qwen3.7-flash, which remains billable.
  • One dating quirk worth recording: for its first days on the Qwen Cloud model-releases changelog, this entry was stamped August 26, 2025 — a vendor typo. It sat at the top of a strictly reverse-chronological log, above entries dated August 24, 2026, while every other surface (Alibaba's own News log, the HuggingFace commit history, same-day press coverage) said 2026. Alibaba has since corrected it: the changelog read August 26, 2026 when I re-checked on September 2, 2026, with every date heading back to March 2026 in strict descending order.
Model
Qwen3.8-27B
Qwen/Qwen3.8-27B · qwen3.8-27b — hosted, Qwen Cloud since Aug 19, 2026
Flagship
Current
Aug 14, 2026
27B dense, Apache 2.0, and the newest open-weights Qwen flagship. Native vision-language — images and video in — on the Qwen3.5 hybrid Gated DeltaNet + gated-attention foundation, 262K context extensible to 1M, thinking on by default with reasoning_effort control. Alibaba positions it at parity with Qwen3.7-Plus, an MoE roughly ten times its size, and runs it on one 24 GB consumer GPU at 4-bit.
  • Weights published August 14, 2026 — the date Alibaba's own QwenLM/Qwen3.8 News log gives for the HuggingFace and ModelScope drops. HuggingFace card: Qwen/Qwen3.8-27B (an FP8 build, Qwen3.8-27B-FP8, landed the same day). Alibaba's write-up is Alibaba Unveils Qwen3.8-27B and Releases Weights of Qwen3.8 Flagship Model (August 17, 2026).
  • 27B fully dense, natively vision-language. 64 layers, hidden dimension 5,120, a 248,320-token vocabulary, and the Qwen3.5 hybrid layout — 16 × (3 × Gated DeltaNet → FFN, then 1 × gated attention → FFN) — trained with multi-token prediction. Context is 262,144 tokens natively, extensible to 1,000,000. Thinking mode is on by default and can be switched off per request; reasoning_effort tunes depth and preserve_thinking carries reasoning context across turns.
  • License: Apache 2.0 — unlike its 2.4T sibling, which shipped two days earlier under a bespoke revenue-share license. The 27B is the size that keeps the Qwen open-weights story permissive.
  • Built to run on one consumer GPU. Alibaba's own pre-launch sizing piece, What It Actually Takes to Run Qwen3.8-27B Locally (August 5, 2026), puts a 27B dense model at ~54 GB of weights at BF16, ~28.6 GB at 8-bit and ~16.8 GB at the mainstream Q4_K_M quant — so 24 GB of VRAM is the practical floor and 16 GB is exactly where it does not fit.
  • Hosted since August 19, 2026 under the model string qwen3.8-27b (marketplace title: “Qwen3.8-Open-Source”), per the Qwen Cloud model-releases changelog. List price on the marketplace page is $0.50 / $3.00 per million input / output tokens (implicit cache reads $0.10), with 1M context, 262K max reasoning tokens, 5M TPM and 5K RPM — read August 21, 2026. The model card had said the hosted version was “coming soon” as recently as August 14.
  • Alibaba reports the model reached the top five most-liked models on HuggingFace within two days of the drop. It supersedes Qwen3.6-27B in the dense open-weights slot; the MoE companion Qwen3.6-35B-A3B has no Qwen3.8 equivalent yet.
Model
Qwen3.8-Max
qwen3.8-max — hosted, Qwen Cloud / Model Studio · Qwen/Qwen3.8-2.4T-A95B — open weights (text-only) since Aug 12, 2026
Flagship
Current
Aug 3, 2026
The Max-line flagship, generally available three weeks after its WAIC preview: 2.4 trillion total parameters with 95 billion active, sparse MoE plus hybrid attention, native vision-language, 1M-token context, hybrid thinking on by default. $2 / $6 per million tokens. On August 12, 2026 the weights shipped as Qwen/Qwen3.8-2.4T-A95B — the first Qwen-Max-class model to go open, and the largest open-weights language model published anywhere — but under a bespoke Qwen3.8-Max License rather than Apache 2.0, and text-only.
  • Previewed July 19, 2026 at the World AI Conference in Shanghai (Alibaba's recap: Agent-Native Innovations at WAIC 2026) and generally available on August 3, 2026. Alibaba's launch posts are Alibaba Unveils Qwen3.8-Max and the technical write-up Qwen3.8-Max: A New Bar for Coding and Cowork; coverage in MarkTechPost, SiliconANGLE, and TNGlobal.
  • 2.4 trillion total parameters, 95 billion active. The preview disclosed only the total; the launch added the active count. Sparse Mixture-of-Experts plus a hybrid attention mechanism, built on the Qwen3.5 architectural foundation. Native vision-language — text, image, and video in, text out — with a 1,000,000-token context window (991K max input, 131K max output, 262K max reasoning tokens) and hybrid thinking mode enabled by default.
  • Generally available and billable. The DashScope model string is qwen3.8-max, served through Qwen Cloud and Alibaba Cloud Model Studio at $2.00 / $6.00 per million input / output tokens (list price re-read August 21, 2026), with implicit cache reads at $0.25. Qwen Cloud's Choose models page orders the text-generation lineup “from most capable to most cost-effective” and puts qwen3.8-max first, ahead of qwen3.7-plus and qwen3.8-flash — Alibaba's own answer to which model is the flagship. The legacy Model Studio recommended-models page, frozen at a July 15 stamp for four weeks after launch, thawed on August 11, 2026 — it listed the model as served in Beijing, Hong Kong, Singapore, Tokyo, Frankfurt, and US (Virginia) — and has kept pace since, carrying a September 4, 2026 stamp and still leading its text-generation section with qwen3.8-max (stamp re-read September 8, 2026).
  • First snapshot refresh: qwen3.8-max-0902, September 2, 2026. Alibaba added a dated snapshot (alias qwen3.8-max-2026-09-02) it describes as stronger on engineering-scale coding and long-horizon autonomous development, on multi-tool agent orchestration, and on chart reasoning / document parsing / multimodal perception, while keeping the 1M context window, thinking mode, and the built-in tool set. It bills at the same $2.00 / $6.00 per million as the bare alias, on its own marketplace page and in Model Studio's pricing table. Alibaba does not say which snapshot the bare qwen3.8-max alias now resolves to — unlike qwen3.7-max, whose billing row still carries an explicit “currently equivalent to qwen3.7-max-2026-05-20” note. Dated snapshots don't get their own rows here; the ship date above stays the August 3 general-availability date.
  • Vendor-cited arena placements: fifth in Text Arena, second in Vision Arena, fourth in Frontend Code Arena at launch. Alibaba's own long-horizon demonstrations are the more distinctive claim — a ~16-day fully autonomous run that produced the open-sourced oh-my-cli agent harness (265 commits / 127 PRs / 151 issues as of July 30, 2026), a five-day research run that reproduced and then beat a published data-selection paper by 2.7 points on AIME24, and a 24-hour entry into Alibaba's Tianchi WWW2025 multimodal-intent contest that finished ahead of 458 of 526 human teams.
  • The open weights landed on August 12, 2026 — two days after the promised window, as Qwen/Qwen3.8-2.4T-A95B (plus an FP8 build) on HuggingFace and ModelScope, per the News log in QwenLM/Qwen3.8. There is no Qwen/Qwen3.8-Max repo — the open checkpoint carries the parameter-count name. Two caveats matter. It is text-only: the released weights are a causal LM (qwen3_5_moe_text), and Alibaba's own card says the hosted qwen3.8-max is “the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools.” The open checkpoint is 262K natively, extensible to ~1,010,000. And it is not Apache 2.0 — see the next bullet.
  • The open checkpoint is also served as its own hosted model string. A day after the weights landed, on August 13, 2026, Alibaba added qwen3.8-2.4t-a95b to Qwen Cloud — so the text-only open build is billable alongside the vision-capable qwen3.8-max, not just downloadable. Its marketplace page carries the same $2.00 / $6.00 per million list price and 1M context, but 131K max reasoning tokens rather than 262K, 5M TPM / 5K RPM rather than 2M / 15K, and only the code_interpreter / web_search / web_extractor built-in tools. Alone among the hosted Qwen3.8 strings it lacks the t2i_search / i2i_search image-search tools, which qwen3.8-max, qwen3.8-flash and qwen3.8-27b all carry (re-read September 8, 2026).
  • A fourth license tier: the Qwen3.8-Max License. The repo's LICENSE grants MIT-shaped rights to use, modify, host, fine-tune, and sell, with two conditions. Products above 100 million monthly active users or $20 million monthly revenue must display the model name prominently in their UI. And any licensee running a “Model as a Service” or “AI Work Assistant” business whose group revenue exceeds $50 million over any twelve consecutive months must obtain a separate commercial license from Qwen — internal use is carved out. HuggingFace read the same shift in its State of Open Models: Summer 2026 (August 14, 2026): the very largest Chinese releases, “Kimi K3 and Qwen 3.8 2.4T,” have started attaching non-commercial restrictions and revenue-share terms while everything below that size band stays Apache or MIT. The companion Qwen3.8-27B, shipped two days later, is plain Apache 2.0.
  • Alibaba shipped the model alongside QwenWork, an all-in-one workplace agent platform that entered public beta in China on the same day and offers Qwen3.8-Max as its Flagship model tier. The July preview had landed two days after Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model; the timing was widely read as a response.
Model
Qwen-Audio-3.0-ASR-Flash
qwen-audio-3.0-asr-flash · -filetrans · -streaming — hosted, Qwen Cloud / Model Studio
Specialized
Current
Jul 30, 2026
Third-generation speech recognition in three hosted variants — batch, file-transcription, and streaming. Seven Chinese dialect groups plus 20+ regional accents, 30 languages, built-in industry lexicons, hotword and context injection. Alibaba reports 95.36% medical-term recall. Hosted-only: no HuggingFace weights, unlike the Apache 2.0 Qwen3-ASR line it succeeds.
  • API availability dated July 30, 2026 on the Qwen Cloud model-releases changelog, with the Chinese-language launch on July 31, 2026 (AIBase, translated title: “Alibaba releases new speech-recognition model, medical vocabulary accuracy exceeds 95%”). This release never got an alibabacloud.com/blog post.
  • Three variants, one lineage. qwen-audio-3.0-asr-flash for non-real-time recognition of clips up to five minutes, qwen-audio-3.0-asr-flash-filetrans for offline file transcription, and qwen-audio-3.0-asr-flash-streaming for real-time streaming.
  • Dialect and language coverage: the seven major Chinese dialect groups (Mandarin, Wu, Xiang, Gan, Hakka, Min, Yue) plus more than 20 regional accents, and 30 languages including Chinese, English, Japanese, and Korean.
  • Domain-vocabulary focus is the pitch: pre-built industry lexicons (healthcare, IT, finance, public figures) plus precompiled and on-the-fly hotwords and context injection, so specialist terms resolve without per-deployment configuration. Alibaba reports 95.36% professional-term recall in medical scenarios and 91.87% in IT on its internal evaluation, and cites a prior first-place finish on Artificial Analysis at a 1.7% error rate. Also adds punctuation prediction, number / date / currency normalization, and classical-Chinese-poetry recognition.
  • Hosted, not open-weights. No Qwen/Qwen-Audio-3.0-ASR repo exists in the HuggingFace Qwen org (verified August 21, 2026) — the same closed pattern as Qwen-Audio-3.0-TTS, Qwen-Audio-3.0-Realtime, and Qwen-Image-3.0. The Apache 2.0 Qwen3-ASR family it succeeds remains the line's latest open-weights speech-recognition release.
Model
Qwen-Image-3.0
qwen-image-3.0-pro · qwen-image-3.0 — hosted, Qwen Cloud / Model Studio (since Aug 4, 2026)
Specialized
Current
Jul 21, 2026
Third-generation image model pitched at information-dense working images: prompts up to 4,500 tokens, text legible at ~10 px, native rendering in 12 languages, live-data pulls. Launched chat-only; API model strings followed on August 4, 2026. Still closed — no weights, license, benchmark table, or technical report — a break from the open-weights precedent of Qwen-Image 1.0 / 2.0.
  • Released July 21, 2026; the announcement is at qwen.ai/blog, with the Alibaba Cloud Community write-up Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge (July 22, 2026) and coverage in Unite.AI.
  • Three headline capabilities, all vendor-demonstrated. Long instructions — up to 4,500 tokens (up from ~1,000 in Qwen-Image-2.0) composed into one information-dense image in a single pass; Alibaba's centerpiece is a 3×3 grid of unrelated infographics generated from a single 3,700-token instruction. Fine detail — text rendered legibly down to ~10 pixels (full academic-paper pages with equations, simulated newspaper front pages). And “world knowledge” — native text rendering across 12 languages, interface reproduction, and live-data pulls such as a weather graphic for a specific city and date.
  • Positioned at content-production work — newspaper layouts, short-drama storyboards, UI mockups, e-commerce imagery — plus editing tasks like adding handwritten annotations and restoring a damaged traditional painting.
  • Closed and hosted-only. The launch shipped no benchmark table, parameter count, license text, downloadable weights, or technical report — a departure from Qwen-Image 1.0 (August 2025, Apache 2.0 with a same-day technical report) and Qwen-Image-2.0 (technical report at arXiv 2605.10730). No Qwen/Qwen-Image-3.0 repo exists in the HuggingFace Qwen org (re-verified August 11, 2026).
  • API access followed two weeks later. The model launched on Qwen Chat only (chat.qwen.ai); the Qwen Cloud changelog dates the API strings to August 4, 2026 — qwen-image-3.0-pro for the long-text, dense-layout tier and qwen-image-3.0 as the standard quality-and-speed variant. Both bill input and output images separately and price the output by resolution tier: Pro at $0.04 (1k) / $0.075 (2k) per image, standard at $0.03 either way, plus $0.003 per input image (list prices read August 11, 2026). qwen-image-3.0-pro is now the recommended image-generation model on Qwen Cloud's Choose models page, displacing qwen-image-2.0-pro.
  • Context for the claims: Alibaba's own Qwen-Image-Bench (arXiv 2605.28091) placed the prior flagship, Qwen-Image-2.0 Pro, fifth overall behind OpenAI's GPT Image 2 / GPT Image 1.5 and Google's Nano Banana models. No independent evaluation of 3.0 exists yet; this page will add weights, license, and API strings if Alibaba publishes them.
Model
Qwen-Audio-3.0-TTS
qwen-audio-3.0-tts-flash · qwen-audio-3.0-tts-plus — hosted, Qwen Cloud / Model Studio
Specialized
Current
Jul 20, 2026
Text-to-speech across 16 languages and 20 Chinese dialect regions, in two hosted tiers — Flash (~300 ms first-packet latency) and Plus (quality-tuned). Natural-language style control plus 86 inline tags. Alibaba says Plus ranks first on the Artificial Analysis TTS leaderboard. Hosted-only: no HuggingFace weights.
  • Released July 20, 2026; the Alibaba Cloud Community write-up Qwen-Audio-3.0-TTS: More Multilingual, Easier to Direct posted July 21, 2026; technical page at funaudiollm.github.io; coverage in MarkTechPost. The DashScope model strings were already live in the Model Studio real-time speech synthesis docs at their July 14, 2026 refresh.
  • Two hosted tiers from one lineage — qwen-audio-3.0-tts-flash is tuned for real-time interaction at a 300 ms-level first-packet latency; qwen-audio-3.0-tts-plus is tuned for naturalness and timbre fidelity.
  • 16 languages (Arabic, Chinese, English, French, German, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Tagalog, Thai, Vietnamese), seven of them newly added, plus 20 Chinese dialect regions. Alibaba reports best WER/CER in 10 of the 16 and an average speaker-similarity of 82.75 for Plus.
  • Two control surfaces. Free-form natural-language instructions steer role, emotion, style, rate, timbre, and accent; 86 new fine-grained inline tags ([gasp], [giggles], [angry]) give phrase- and word-level control over non-verbal detail.
  • Architecture per the technical page: a 12.5 Hz low-frame-rate speech tokenizer with a five-stage progressive training pipeline (independent LM + FM pretraining, joint training with data annealing, LM RL, FM robustness training, FM RL). One-pass long-form synthesis up to three minutes; speech enhancement is built into the voice-cloning path for noisy or reverberant reference clips.
  • Hosted, not open-weights. Shipped as a Model Studio / DashScope API only — no Qwen/Qwen-Audio-3.0-TTS repo exists in the HuggingFace Qwen org and no matching ModelScope page exists either (both re-verified August 11, 2026), unlike the Apache 2.0 Qwen3-TTS line it succeeds. Alibaba lists a curated preset voice library and 48 kHz output as "coming soon."
Model
Qwen-Audio-3.0-Realtime
qwen-audio-3.0-realtime-flash · qwen-audio-3.0-realtime-plus — hosted, Model Studio; Qwen Cloud since Aug 10, 2026
Specialized
Current
Jul 15, 2026
Native speech-to-speech — full-duplex WebSocket voice conversation with semantic turn detection, function calling, and cloned voices, in Flash and Plus tiers. Artificial Analysis ranked Plus first on its Speech-to-Speech Index (84.1%) on July 28, ahead of OpenAI's GPT-Realtime-2.1 High. Hosted-only: no HuggingFace weights.
  • Announced July 15, 2026 in Alibaba's Chinese-language release post “Qwen-Audio-3.0-Realtime: it can chat, and it can get things done” (Zhihu; title translated). Model Studio's Qwen-Audio real-time voice model guide and its speech-to-speech model selector both carried the two model strings at their July 14, 2026 refresh. English coverage arrived with the benchmark result on July 28–30 — Neowin and BetaNews.
  • Native speech-to-speech, not a pipeline. One model takes audio in and returns speech out over a full-duplex WebSocket, so tone and emotion survive the round trip instead of being flattened by an ASR → LLM → TTS chain. Audio is PCM 16 kHz mono in, 24 kHz mono out; the session holds up to 50 turns or 300 seconds of cumulative audio before older history is dropped.
  • Three turn-taking modes — acoustic VAD (server_vad), semantic turn detection (smart_turn, which weighs meaning as well as sound so filler like “uh” or “hmm” doesn't end a turn), and manual push-to-talk. Model Studio's speech-to-speech guide names qwen-audio-3.0-realtime-plus as the recommended model for semantic-VAD voice assistants and customer service with function-calling support; both tiers accept cloned voices minted through the Voice Cloning API.
  • First on an independent speech benchmark. Artificial Analysis placed Qwen-Audio-3.0-Realtime-Plus at 84.1% on its Speech-to-Speech Index on July 28, 2026 — ahead of OpenAI's GPT-Realtime-2.1 High at 79.1%, and first on all three equally-weighted components (Big Bench Audio for speech reasoning, Full Duplex Bench for conversational dynamics, Tau Voice for agentic performance). The tradeoff is latency: roughly 4 seconds to first audio against ~1.14 seconds for GPT-Realtime-2 High.
  • Two tiers, hosted only. qwen-audio-3.0-realtime-plus and qwen-audio-3.0-realtime-flash bill text and audio tokens at separate rates. The pair reached the international Qwen Cloud platform on August 10, 2026, nearly four weeks after the China launch, where Plus lists at $0.80 text / $6.40 audio per million input tokens and $24.00 per million output tokens for combined text-and-audio output, on a 40K context (list prices read August 11, 2026). No Qwen/Qwen-Audio-3.0-Realtime repo exists in the HuggingFace Qwen org and no matching ModelScope page exists either (both re-verified August 11, 2026) — one of four closed specialized debuts in the July 2026 wave, alongside Qwen-Audio-3.0-TTS, Qwen-Image-3.0, and Qwen-Audio-3.0-ASR-Flash.
Model
Qwen-AgentWorld
Qwen/Qwen-AgentWorld-35B-A3B — open weights; the larger 397B-A17B is benchmark-only, no public repo
Specialized
Current
Jun 25, 2026
First Qwen "language world model" — predicts the next environment state (terminal output, API response, DOM) given an agent's action, across seven agent domains in one model. 35B-A3B open weights (Apache 2.0) built on Qwen3.5; 397B-A17B tops AgentWorldBench over GPT-5.4, Claude Opus 4.8, and Gemini 3.1 Pro.
  • Released June 25, 2026; the announcement is on qwen.ai/blog and the Alibaba Cloud Community blog; technical report at arXiv 2606.24597; GitHub: QwenLM/Qwen-AgentWorld; coverage in VentureBeat.
  • Qwen's first "language world model." Rather than deciding the next action, it predicts the next observation — what the terminal, browser, or Android screen returns given the agent's action and interaction history. Environment modeling is the training objective from continual pre-training onward (CPT → SFT → RL), not a post-hoc adaptation on a general-purpose LLM.
  • Seven agent domains in one model — text-based (MCP, Search, Terminal, SWE) and GUI-based (Web, OS, Android). For the GUI domains, observations are renderable code (accessibility-tree XML, HTML, UI-hierarchy markup) rather than pixels, enabling text-only world modeling of visual environments.
  • Two scales. The open-weights Qwen/Qwen-AgentWorld-35B-A3B (35B total / 3B active MoE, 256K context, built on Qwen3.5-35B-A3B-Base) plus the larger Qwen-AgentWorld-397B-A17B, which tops AgentWorldBench with an overall 58.71, ahead of GPT-5.4 (58.25), Claude Opus 4.8 (56.59), and Gemini 3.1 Pro (54.57) per Alibaba's benchmarks. Only the 35B-A3B has a published repo: Qwen/Qwen-AgentWorld-397B-A17B does not exist on HuggingFace or ModelScope (both probed August 11, 2026), so the 397B figures are Alibaba-reported and not independently reproducible.
  • License: Apache 2.0 (confirmed on the HuggingFace card: license: apache-2.0). Weights on HuggingFace and ModelScope, alongside the AgentWorldBench evaluation set (seven domains, ground-truth observations from real environments).
  • Positioned as a complementary axis to real-environment training: as a decoupled simulator it enables controllable agentic-RL rollouts that surpass RL against live environments, and as a unified agent foundation model its world-modeling warm-up transfers to multi-turn tool-calling tasks without task-specific fine-tuning.
Model
Qwen-Robot Suite
Qwen-RobotManip, Qwen-RobotNav, Qwen-RobotWorld — enterprise pilot
Specialized
Current
Jun 16, 2026
Qwen's first robotics / physical-AI foundation models. Three models — manipulation (VLA), navigation (VLN), and a video world model — built on the Qwen3-VL and Qwen3.5-VL stack. Tops the RoboChallenge real-robot benchmark. In pilot with select enterprise customers; weights / license not yet disclosed.
  • Announced June 16, 2026 on qwen.ai; the launch write-up is Alibaba's Entering the Physical AI Era: Introducing the Qwen-Robot Suite, with per-model posts for RobotManip, RobotNav, and RobotWorld. Coverage in eWeek and South China Morning Post.
  • Qwen's first move into embodied / physical AI — the company's stated goal is to bridge vision-language understanding and physical control ("seeing is not acting"). The general-purpose Qwen models compose with the robotics models as specialized tools, so an agentic workflow can use a Qwen flagship as the planner and a robot model as the real-time execution tool.
  • Qwen-RobotManip — a generalizable Vision-Language-Action (VLA) model built on Qwen3.5-4B VL, trained on 38,000+ hours of purely open-source data (robotics repositories, human-manipulation video, synthesized human-to-robot datasets). Delivers a 3× improvement over prior SOTA in cross-embodiment transfer; ranked first (codenames Lira / Atlas) on the RoboChallenge Table30 generalist track.
  • Qwen-RobotNav — a Vision-Language-Navigation (VLN) model built on Qwen3-VL, trained on 15.6 million trajectory-planning + vision-language-reasoning samples. Serves as both a scalable navigation engine and a unified interface for agentic navigation, including embodied question answering.
  • Qwen-RobotWorld — a video world model that predicts physically-grounded future visual trajectories from current observations. Trained on 8.6 million video-text pairs (200M+ frames) across 20+ embodiment types and 500 action categories; can synthesize robot training video and simulate trajectories before execution.
  • Pilot, not general release. Alibaba says the suite has entered pilot testing with selected Alibaba Cloud enterprise robotics customers; as of June 2026 no model weights, API model strings, or license terms have been publicly disclosed — so the row carries no HuggingFace id or DashScope string yet. This page will add those when Alibaba publishes them.
Model
Qwen3.7-Plus
qwen3.7-plus — closed-weights, Bailian / Model Studio (DashScope)
Flagship
Current
May 31, 2026
Multimodal agent flagship: vision + language (image and video understanding), deep reasoning, tool invocation, and autonomous iteration. The multimodal sibling to the text-only Qwen3.7-Max.
  • Announced on qwen.ai/research on May 31, 2026; coverage in MarkTechPost (June 2, 2026). Available via Alibaba Cloud Bailian / Model Studio (DashScope model string: qwen3.7-plus). Listed as the recommended multimodal model on the Alibaba Cloud Bailian platform.
  • Multimodal vision + language — understands images and video alongside text; visual understanding, not generation. The multimodal complement to the text-only Qwen3.7-Max; together they form the Qwen3.7 generation announced at the May 20, 2026 Alibaba Cloud Summit.
  • Five agentic capabilities: deep reasoning (step-by-step problem solving), self-programming (writes and revises its own code), tool invocation (calls external functions or APIs), verification and testing (runs outputs and checks results), and autonomous iteration (loops until the task is complete). Alibaba positions the model as a step in multimodal hybrid agent technology.
  • GUI agent integration on the Bailian platform: can operate graphical interfaces via screenshot understanding, handle browser-based tasks, and execute shell commands — the orchestration logic is baked into the model rather than the agent framework.
  • Vision Arena (LM Arena): the Qwen3.7-Plus-Preview ranked #16 overall, placing Alibaba as the #5 lab in vision among all competing AI labs at release time.
  • Closed-weights, proprietary. Available exclusively through Alibaba Cloud Bailian / Model Studio (DashScope) — no HuggingFace open-weights release announced as of June 2026. Alibaba's Bailian platform pairs the model with an Agentic RL mechanism that uses real-world execution feedback to refine accuracy, alongside built-in safety guardrails for autonomous tool use.
  • A cost-tier sibling shipped July 25, 2026: qwen3.7-flash (plus the qwen3.7-flash-2026-07-15 snapshot), a native vision-language Flash model that Alibaba describes as a comprehensive upgrade over 3.6-Flash on multimodal understanding, Search / CI agent execution, and multimodal coding — per the Qwen Cloud model-releases changelog. Priced from $0.03 / $0.13 per million tokens at short context. Following the same convention as the Qwen3.6 tier row below, the Flash SKU is recorded here rather than as its own row.
  • Still the recommended middle tier after Qwen3.8-Max shipped: Qwen Cloud's capability-ordered Choose models page lists qwen3.8-max, then qwen3.7-plus (“balanced performance, speed, and cost”), then qwen3.8-flash, which took the cost tier from qwen3.7-flash on August 26 — and there is still no Qwen3.8-Plus: the marketplace returns 404 for qwen3.8-plus and no such string appears in the Model Studio billing table. Pricing is $0.40 / $1.60 per million tokens up to 256K of input, $1.20 / $4.80 above that, currently carrying a limited-time 20% discount (re-read September 8, 2026).
Model
Qwen3.7-Max
qwen3.7-max — closed-weights, DashScope + OpenRouter / Together / Qubrid
Flagship
Available
May 20, 2026
Reasoning-agent flagship. 1M-token context, native extended-thinking. Beats Claude Opus 4.6 on Terminal-Bench 2.0 / SWE-Bench Pro / MCP-Atlas. $2.50 / $7.50 per million tokens — now more expensive than the Qwen3.8-Max that superseded it at the top of the Max tier on August 3, 2026.
  • Announced at the Alibaba Cloud Summit in Hangzhou on May 20, 2026; commercial API live on Alibaba Cloud Model Studio one day earlier (May 19). Cross-listings on OpenRouter, Together AI, and Qubrid AI by day zero. The launch write-up is Alibaba's Qwen3.7: The Agent Frontier blog post, with coverage in MarkTechPost and TechNode (May 21, 2026).
  • 1,000,000-token context window and native extended-thinking mode — the first Qwen Max-tier model with both as defaults.
  • Vendor-stated benchmarks at launch: SWE-Bench Pro 60.6, Terminal-Bench 2.0 69.7, GPQA Diamond 92.4, Artificial Analysis Intelligence Index 56.6 — positioned ahead of DeepSeek V4-Pro and Claude Opus 4.6 on agentic-coding evaluations.
  • Closed-weights, proprietary — like the rest of the Max line, no HuggingFace release. The page's licensing card now lists three tiers: Apache 2.0 (the open Qwen3.6 family), Tongyi-Qianwen License (the older specialized weights), Proprietary (the Max line).
  • Pricing $2.50 / $7.50 per million input / output tokens — roughly half of Claude Opus 4.7's rate card — re-confirmed on Qwen Cloud's pricing page (list price, read August 11, 2026). The bare qwen3.7-max alias currently resolves to the qwen3.7-max-2026-05-20 snapshot; later snapshots (-2026-06-08) and a US-region variant (qwen3.7-max-us) are billed separately. Vendor demos emphasized 35-hour autonomous-agent runs without performance degradation.
  • Superseded at the top of the Max tier by Qwen3.8-Max on August 3, 2026, which is both more capable on Alibaba's own ordering and cheaper ($2.00 / $6.00 versus $2.50 / $7.50). qwen3.7-max is still served and billed, and its dated snapshots and the qwen3.7-max-us region variant remain addressable — but it no longer heads Qwen Cloud's capability-ordered model list.
Model
Qwen3.6-27B
Qwen/Qwen3.6-27B
Flagship
Available
Apr 22, 2026
27B dense, Apache 2.0. Hybrid Gated DeltaNet + self-attention. Thinking Preservation. 262K context (extensible to 1M). Beats 397B MoE on coding. Superseded in the dense open-weights slot by Qwen3.8-27B on August 14, 2026.
  • Released April 22, 2026; the announcement is at qwen.ai/blog; HuggingFace card: Qwen/Qwen3.6-27B. Coverage in MarkTechPost and Simon Willison.
  • 27B fully dense — all parameters active on every inference pass, simplifying deployment vs. the MoE pattern most peer flagships use. The first fully dense flagship in the Qwen3.6 family.
  • Hybrid Gated DeltaNet + self-attention architecture; introduces Thinking Preservation, a mechanism that retains reasoning traces across conversation history to reduce redundant token generation in multi-turn agent workflows.
  • Default 262,144-token context, extensible to 1,010,000 tokens. Supports both multimodal thinking and non-thinking modes.
  • License: Apache 2.0. Per Alibaba's release post, scores 77.2 on SWE-bench Verified and 53.5 on SWE-bench Pro, beating the much larger Qwen3.5-397B-A17B on agentic-coding benchmarks at a 14× smaller active-parameter footprint.
  • Designed to run on a single consumer GPU; community testing reportedly hit ~80 tokens/second on a single RTX 5090 with a 218K-token context window.
  • Demoted to Available on August 21, 2026. Qwen3.8-27B (August 14, 2026) is the named successor — the same 27B dense slot, the same Apache 2.0 license, built on this model per Alibaba's own framing (“builds upon the 3.6-27B version”). The weights stay on HuggingFace and the model remains a reasonable choice for anyone already deployed on it.
Model
Qwen3.6-35B-A3B
Qwen/Qwen3.6-35B-A3B
Flagship
Current
Apr 16, 2026
First open-weights Qwen3.6 release: 35B total / 3B active MoE. Apache 2.0. The MoE companion to the dense Qwen3.6-27B shipped six days later.
  • Released April 16, 2026 on HuggingFace and ModelScope; HuggingFace card at Qwen/Qwen3.6-35B-A3B.
  • 35B-total / 3B-active MoE — the small-active-parameter sibling that pairs with the fully dense Qwen3.6-27B in the open-weights Qwen3.6 wave.
  • License: Apache 2.0. The first publicly-released Qwen3.6-family open-weights model, preceding the dense 27B by six days and the broader open-weights wave.
  • Designed for cost-efficient agentic-coding workloads where the active-parameter budget matters more than total scale; GGUF community-quantized builds followed shortly on Ollama and llama.cpp.
Model
Qwen 3.6-Max-Preview + Qwen 3.6-Plus + Qwen 3.6-Flash
qwen3.6-max-preview, qwen3.6-plus, qwen3.6-flash — closed-weights, DashScope
Flagship
Legacy
Apr 2, 2026
Proprietary closed-weights tier. OpenAI- and Anthropic-compatible API. Three SKUs from most capable to most cost-effective: Max-Preview, Plus, Flash. Superseded by the Qwen3.7 line and then by Qwen3.8-Max; all three SKUs are still billable, but none appears on Qwen Cloud's recommended-models list (read August 11, 2026).
  • Released April 2, 2026; coverage in Caixin Global and Alibaba Cloud Community.
  • Proprietary, closed-weights, DashScope-only. Per Alibaba's framing, the most powerful Qwen model shipped to date; tops six major coding benchmarks and posts gains on world-knowledge and instruction-following over Qwen 3.5-Plus.
  • Dual-API compatibility — the API is compatible with both OpenAI and Anthropic specifications, so existing pipelines can be re-pointed with minimal changes.
  • Available via Alibaba Cloud Model Studio. Pairs with the open-weights Qwen3.6-27B (above) as parallel tracks of the same flagship product slot — the open-weights track for self-host, the proprietary track for hosted commercial use.
Model
Qwen3.5-Omni
qwen3.5-omni-plus, qwen3.5-omni-plus-realtime — DashScope / Model Studio
Specialized
Current
Mar 30, 2026
Native omni-modal: text + image + audio + video in, real-time speech out. Plus / Flash / Light tiers, 256K-token context. 113 speech-recognition languages. Matches Gemini-3.1 Pro on audio understanding.
  • Released March 30, 2026; coverage in MarkTechPost and eWeek. Listed on the Alibaba Cloud Model Studio recommended omni / speech lists; DashScope strings qwen3.5-omni-plus and qwen3.5-omni-plus-realtime.
  • Native end-to-end omni-modal — processes text, images, audio, and video in a single pass and streams responses in text and natural speech in real time. The successor to Qwen3-Omni (September 2025) and Qwen2.5-Omni-7B (March 2025).
  • Plus / Flash / Light tiers for different capability-versus-cost budgets; the flagship qwen3.5-omni-plus carries a 256K-token context — enough for roughly ten hours of audio or ~400 seconds of 720p video at 1 fps.
  • 113 speech-recognition languages and dialects (up from 19 in Qwen3-Omni) with speech generation across 36 languages; per Alibaba's benchmarks it outperforms Gemini-3.1 Pro on general audio understanding, reasoning, recognition, and translation, and matches it on audio-video comprehension.
Model
Qwen3.5 family + Plus
Qwen/Qwen3.5-{0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B, 397B-A17B}, qwen3.5-plus (DashScope)
Flagship
Available
Feb 16, 2026
Multimodal MoE family. Open weights up to 397B. Native text + image + video. 201 languages. Hybrid Gated DeltaNet + sparse MoE.
  • Qwen3.5 (Qwen/Qwen3.5-397B-A17B) and Qwen3.5-Plus released February 16, 2026; the smaller open-weights sizes (27B, 35B-A3B, 122B-A10B) followed February 24, 2026. Coverage in CNBC and SiliconANGLE.
  • A small-model wave closed the family out at the end of February — Qwen/Qwen3.5-9B and Qwen3.5-4B (February 27, 2026), then Qwen3.5-2B and Qwen3.5-0.8B (February 28), each with a matching -Base repo, all Apache 2.0 (repo creation dates and license fields read from the HuggingFace API, July 29, 2026). These are the sizes that put Qwen3.5 on laptops and phones, and they are the most-downloaded models in the family. It was days after this wave that Qwen's technical lead resigned — see People who shaped Qwen.
  • First Qwen with native multimodality across text + image + video in a single architecture, with early-fusion training on trillions of multimodal tokens; outperformed Qwen3-VL on broad reasoning / coding / agents benchmarks per Alibaba's release post.
  • Hybrid Gated DeltaNet + sparse Mixture-of-Experts architecture — the open-weights flagship at 397B total parameters; the smaller MoE variants (35B-A3B, 122B-A10B) and the dense 27B target progressively smaller deployment footprints.
  • 201 languages and dialects — expanded from Qwen3's 119 languages, the broadest language coverage of any frontier-AI line.
  • Open weights under Apache 2.0; Qwen3.5-Plus is the proprietary hosted variant on DashScope / Model Studio.
  • Superseded in the open-weights flagship slot two months later by the Qwen3.6 generation — Qwen3.6-35B-A3B (April 16, 2026) and Qwen3.6-27B (April 22, 2026) — and since by Qwen3.8-27B. The weights are still served.
Model
Qwen-Image-2.0
qwen-image-2.0-pro (DashScope) · open weights via QwenLM/Qwen-Image
Specialized
Available
Feb 10, 2026
Unified text-to-image generation + editing in one 7B model. Native 2K resolution, professional Chinese + English typography. #1 on AI Arena for both generation and editing at release. Still the latest open-weights image release, but no longer the recommended image API — Qwen-Image-3.0 took that slot when its API strings shipped in August 2026.
  • Released February 10, 2026; the announcement is on the Alibaba Cloud Community blog, with coverage in GIGAZINE. Open weights via the QwenLM/Qwen-Image GitHub; the hosted Pro tier is qwen-image-2.0-pro on DashScope / Model Studio.
  • Unified generation + editing — folds text-to-image generation and image editing into a single 7B architecture. Superseded Qwen-Image-2512 (now Legacy) as the recommended image model. Superseded in turn by Qwen-Image-3.0 (July 21, 2026), which took over as the recommended image-generation API on August 4, 2026 when its qwen-image-3.0-pro / qwen-image-3.0 strings shipped. qwen-image-2.0-pro is still served and billed ($0.075 per image, read August 11, 2026), and Qwen-Image-2.0 remains the line's latest open-weights release.
  • Native 2K resolution with pixel-level layout control for complex multi-element compositions — infographics, posters, and slide pages with accurate Chinese-and-English typography from instructions up to ~1K tokens.
  • Held the #1 position on AI Arena (blind human-preference evaluation) in both the text-to-image and image-editing categories at release.
Model
Qwen3-Coder-Next
Qwen/Qwen3-Coder-Next — open weights · qwen3-coder-next — hosted, Qwen Cloud / Model Studio
Specialized
Current
Feb 4, 2026
Open-weights coding agent. 80B total / 3B active (Qwen3-Next architecture). Apache 2.0. 70%+ SWE-Bench Verified. 256K context, 10× throughput vs. dense 32B. Still the newest release in the Coder slot — no successor has shipped.
  • Released February 4, 2026; the announcement is at qwen.ai/blog; HuggingFace card: Qwen/Qwen3-Coder-Next; coverage in MarkTechPost and VentureBeat.
  • Built on Qwen3-Next-80B-A3B-Base — 80B total parameters with only 3B activated per inference step (ultra-sparse MoE + hybrid Gated DeltaNet + Gated Attention architecture). Specifically tuned for coding-agent workloads: long-horizon reasoning, complex tool use, and recovery from execution failures.
  • 256K native context, extensible to 1M tokens. Non-thinking-only mode (no think blocks in output) — optimized for deterministic agent pipelines rather than exploratory chain-of-thought.
  • Per Alibaba's release post: 70%+ on SWE-Bench Verified (SWE-Agent scaffold) and 44.3% on SWE-Bench Pro — competitive with the much larger Qwen3-Coder-480B on agentic-coding benchmarks at a dramatically lower active-parameter footprint.
  • License: Apache 2.0. More than 10× higher throughput than Qwen3-32B on context lengths exceeding 32K tokens, enabling cost-efficient deployment for repository-scale tasks.
  • Also hosted, not only downloadable. The model string qwen3-coder-next is billable on Qwen Cloud (marketplace title “Qwen3-Open-Source”) at $0.30 / $1.50 per million input / output tokens for requests up to 32K of input, with a 262K context, 204K max input, 65K max output, 1M TPM and 600 RPM — read September 2, 2026 on the marketplace page, and priced the same in Model Studio's billing table alongside the older qwen3-coder-plus and qwen3-coder-flash strings. It never appeared in the Qwen Cloud release log, which is why it is easy to miss.
  • Why this row is still Current. Every other superseded Specialized row here has an identifiable successor; the Coder slot does not. No Qwen coding model has shipped since — the QwenLM/Qwen3-Coder repo has had no release since March 2026, the Qwen Cloud release log carries no coder entry of any kind, and coding is instead pitched as a headline capability of the general flagships (Qwen3.8-Max, Qwen3.6-27B). The Qwen Cloud marketplace also returns 404 for qwen3.8-coder and qwen3.8-coder-plus, while qwen3-coder-next still resolves. So this remains the latest coding release in the line (re-checked September 8, 2026).
Model
Qwen3-ASR + Qwen3-ForcedAligner
Qwen/Qwen3-ASR-{0.6B, 1.7B} · Qwen/Qwen3-ForcedAligner-0.6B
Specialized
Available
Jan 29, 2026
First Qwen speech-recognition line: all-in-one ASR + language ID for 52 languages and dialects, built on Qwen3-Omni. Apache 2.0. Qwen3-ASR-1.7B is SOTA among open-source ASR (beats Whisper-large-v3, competitive with GPT-4o-Transcribe); the non-autoregressive Qwen3-ForcedAligner-0.6B adds timestamp alignment.
  • Released January 29, 2026; the announcement is on qwen.ai/blog and the Alibaba Cloud Community blog; technical report at arXiv 2601.21337; GitHub: QwenLM/Qwen3-ASR; HuggingFace collection: Qwen3-ASR.
  • Qwen's first dedicated speech-recognition family. Two all-in-one models — Qwen/Qwen3-ASR-1.7B and Qwen/Qwen3-ASR-0.6B — that perform language identification and ASR for 52 languages and dialects (30 languages plus 22 Chinese dialects), from a single unified model built on the Qwen3-Omni foundation.
  • Streaming + offline in one model, robust to noise, background music, and singing voice. Per the technical report, Qwen3-ASR-1.7B is state-of-the-art among open-source ASR models and competitive with the strongest proprietary APIs (GPT-4o-Transcribe, Gemini-2.5-Pro, Doubao-ASR), while the 0.6B version hits an accuracy-efficiency sweet spot at high concurrency.
  • Qwen3-ForcedAligner-0.6B — a novel non-autoregressive forced-alignment model that predicts word- / character-level timestamps for text–speech pairs in 11 languages over clips up to 5 minutes, reportedly surpassing WhisperX and NeMo Forced Aligner on long-form audio.
  • License: Apache 2.0 across the family (confirmed on the GitHub repo and HuggingFace cards). Also served via the DashScope real-time and file-transcription speech-recognition APIs on Alibaba Cloud Model Studio.
  • Superseded as the recommended ASR API by the hosted Qwen-Audio-3.0-ASR-Flash family (July 30, 2026), which Qwen Cloud now lists first for speech recognition. Qwen3-ASR remains the line's latest open-weights speech-recognition release — the successor generation shipped closed — and the qwen3-asr-flash hosted endpoint is still billed at $0.000035 per second of audio (read August 11, 2026).
Model
Qwen-Image-2512
Qwen/Qwen-Image-2512
Specialized
Legacy
Dec 2025
Open-weights text-to-image. Apache 2.0. Positioned as an open alternative to Google's Imagen for native image generation.
  • Released December 2025; coverage in Open Source For U.
  • Open-weights text-to-image model, Apache 2.0, distributed via the HuggingFace Qwen org.
  • Positioned by Alibaba as an open alternative to Google's Imagen 4 for native image generation; complements the multimodal capabilities of Qwen3-VL / Qwen3.5 / Qwen2.5-Omni in the same product family.
Model
Qwen3-Max
qwen3-max — closed-weights, DashScope only
Flagship
Legacy
Sep 2025
Trillion-parameter MoE flagship. Proprietary. 36T training tokens. Reportedly top-3 on LMArena alongside GPT-5-Chat at launch.
  • Released September 2025 as the company's first trillion-parameter Qwen flagship; coverage in eWeek.
  • ~1 trillion total parameters in a Mixture-of-Experts architecture; trained on 36 trillion tokens, twice Qwen2.5's pretraining corpus. Per Alibaba, the training run completed without loss spikes — an unusually stable trillion-parameter run.
  • Proprietary, closed-weights, DashScope-only. Per Alibaba's release post, Qwen3-Max-Instruct ranked consistently in the global top three on the LMArena text leaderboard at launch, surpassing GPT-5-Chat.
  • Superseded by the Qwen 3.6-Max-Preview proprietary preview in April 2026; status is Legacy on the “no longer the recommended Max-line model” reading, though the API remained served on DashScope.
Model
Qwen3-VL family
Qwen/Qwen3-VL-{2B, 4B, 8B, 32B, 30B-A3B, 235B-A22B}-Instruct
Specialized
Available
Sep 23, 2025
Vision-language family built on Qwen3. Dense (2B–32B) and MoE (30B-A3B / 235B-A22B) sizes. Instruct + Thinking editions. Apache 2.0. 256K context, image + video.
  • Qwen3-VL-235B-A22B released September 23, 2025; smaller sizes (30B-A3B: October 4; 4B/8B: October 15; 2B/32B: October 21) followed through October 2025. Technical report at arXiv 2511.21631; GitHub: QwenLM/Qwen3-VL; coverage by Simon Willison.
  • Six sizes — four dense (2B, 4B, 8B, 32B) and two MoE (30B-A3B and the flagship 235B-A22B) — each available in Instruct and Thinking editions, mirroring the Qwen3 flagship's hybrid-mode recipe at the vision-language layer.
  • 256K native interleaved context supporting text, images, and video in a single pass; dynamic-resolution visual encoder handles variable-aspect-ratio images and longer video clips.
  • Supersedes the Qwen2.5-VL family (January 2025) as the recommended open-weights vision-language line; the Qwen3.5 flagship's native multimodality later subsumed this specialized track for most use cases.
  • License: Apache 2.0 across all sizes.
Model
Qwen3-Omni
Qwen/Qwen3-Omni-30B-A3B-{Instruct, Thinking}
Specialized
Available
Sep 22, 2025
End-to-end omni-modal: text + image + audio + video in, text + real-time speech out. 30B-A3B MoE, Thinker–Talker architecture. Apache 2.0. SOTA on 22 of 36 audio/video benchmarks.
  • Released September 22, 2025; GitHub: QwenLM/Qwen3-Omni; HuggingFace collection: Qwen3-Omni; technical report: arXiv 2509.17765; API docs: DashScope (Qwen-Omni). Reached #1 on HuggingFace Trending on September 26, 2025.
  • Natively end-to-end omni-modal — processes text, images, audio, and video in a single architecture and delivers real-time streaming responses in both text and natural speech. The successor to Qwen2.5-Omni-7B (March 2025), with a fully redesigned architecture and significantly expanded scale.
  • Novel Thinker–Talker architecture: the Thinker component handles reasoning (chain-of-thought in both Instruct and Thinking modes); the Talker handles real-time speech synthesis with a multi-codebook design that drives latency to a minimum. The MoE base (30B total / 3B active) provides inference efficiency. A separate Thinking-only edition (Qwen/Qwen3-Omni-30B-A3B-Thinking) ships without the Talker for text-only reasoning pipelines.
  • Per the technical report: SOTA on 22 of 36 audio/video benchmarks; open-source SOTA on 32 of 36. Audio speech recognition and voice-conversation performance is comparable to Gemini 2.5 Pro per vendor-stated benchmarks. Supports 119 text languages, 19 speech-input languages, and 10 speech-output languages.
  • License: Apache 2.0 (confirmed on HuggingFace model card: license_name: apache-2.0). Available on HuggingFace, ModelScope, Qwen Chat (chat.qwen.ai), and via the DashScope API. A downstream fine-tune, Qwen3-Omni-30B-A3B-Captioner, ships as a detailed audio-captioning model built on the Instruct base.
  • Superseded as the recommended open-weights end-to-end omni model by the native multimodality introduced in Qwen3.5 (February 2026).
Model
Qwen3-Next-80B-A3B
Qwen/Qwen3-Next-80B-A3B-Instruct, Qwen/Qwen3-Next-80B-A3B-Thinking
Flagship
Available
Sep 11, 2025
Ultra-sparse MoE. 80B total / 3B active (3.7%). Hybrid Gated DeltaNet + Gated Attention. Apache 2.0. Matches Qwen3-235B-A22B at >10× throughput on long context. 256K native.
  • Released September 11, 2025; announcement at Alibaba Cloud Community; HuggingFace cards: Qwen/Qwen3-Next-80B-A3B-Instruct and Thinking.
  • Novel ultra-sparse MoE architecture: activates only 3B of 80B total parameters per inference step (3.7% activation ratio) via hybrid Gated DeltaNet + Gated Attention, a new attention design combining state-space-model-style recurrence with standard self-attention for ultra-long-context efficiency. The architecture later underpins both Qwen3-Next-80B-A3B (base) and Qwen3-Coder-Next (February 2026).
  • 256K native context, extensible to 1M tokens. Per Alibaba, surpasses Qwen3-235B-A22B on long-context benchmarks at more than 10× higher throughput for context lengths exceeding 32K tokens.
  • Surpasses the dense Qwen3-32B model on standard benchmarks while using less than 10% of its training compute (GPU-hours), demonstrating that ultra-sparse MoE can match large-scale dense models at a fraction of the cost.
  • License: Apache 2.0. Available in Instruct and Thinking editions on HuggingFace, ModelScope, Kaggle, and AWS Bedrock.
  • Superseded in the architecture-preview slot by Qwen3.8-Flash-Next (August 26, 2026), which Alibaba describes as playing “the same role that Qwen3-Next played for Qwen3.5” — this time previewing the Qwen4 architecture. The Gated DeltaNet hybrid introduced here carried Qwen3.5 through Qwen3.8 before being reworked into the Gated DeltaNet + Qwen Sparse Attention pairing of its successor.
Model
Qwen3-Coder
Qwen/Qwen3-Coder-{480B-A35B, 30B-A3B}-Instruct
Specialized
Available
Jul 22, 2025
Open-weights agentic-coding flagship. 480B-A35B MoE. Apache 2.0. Reportedly competitive with Claude Sonnet 4 / GPT-4 on coding tasks.
  • Released July 22, 2025; the announcement is at qwenlm.github.io/blog/qwen3-coder; HuggingFace collection: Qwen3-Coder; coverage in MarkTechPost.
  • Qwen3-Coder-480B-A35B-Instruct — 480B total / 35B active per token MoE, the company's most powerful open-weights coding model. 256K context natively, extrapolated to 1M.
  • Qwen3-Coder-30B-A3B-Instruct — the smaller MoE variant for laptop-scale agentic-coding deployment.
  • License: Apache 2.0. Alibaba announced it as “our most agentic code model to date” and claimed state-of-the-art results among open models on agentic coding, browser-use, and tool-use, “comparable to Claude Sonnet 4” (Qwen3-Coder launch post, July 22, 2025) — vendor numbers, and the comparison named Sonnet only.
  • Superseded in the coding slot by Qwen3-Coder-Next (February 4, 2026), which Alibaba's QwenLM/Qwen3-Coder README now lists alongside the 480B and 30B sizes as part of the same line.
Model
Qwen3 family
Qwen/Qwen3-{0.6B, 1.7B, 4B, 8B, 14B, 32B, 30B-A3B, 235B-A22B}
Flagship
Legacy
Apr 28, 2025
Hybrid Thinking / Non-Thinking modes in a single architecture. 6 dense + 2 MoE sizes. 36T tokens, 119 languages. All Apache 2.0.
  • Released April 28, 2025; the announcement is at qwenlm.github.io/blog/qwen3; coverage in TechCrunch; GitHub: QwenLM/Qwen3.
  • First Qwen with hybrid Thinking / Non-Thinking modes in a single architecture — the model dynamically switches between fast direct responses and chain-of-thought reasoning based on the prompt. Absorbed the QwQ standalone-reasoning track into the Flagship V-series.
  • Six dense sizes (0.6B, 1.7B, 4B, 8B, 14B, 32B) and two MoE sizes (30B-A3B and 235B-A22B) released together. The flagship Qwen3-235B-A22B activates 22B per token from 235B total.
  • Trained on 36 trillion tokens — double Qwen2.5's pretraining corpus — with leading performance on translation and multilingual instruction-following across 119 languages and dialects.
  • License: Apache 2.0 across every size, the broadest fully-permissive Qwen launch to date.
  • Superseded as the recommended open-weights flagship by Qwen3.5 (February 2026) and Qwen3.6-27B (April 2026); status is Legacy.

The Qwen3 hybrid-reasoning era — April 28, 2025. Above this line: every Qwen flagship from Qwen3 onward ships with hybrid Thinking / Non-Thinking modes in a single architecture, absorbing the QwQ standalone-reasoning track. Below: the pre-hybrid lineage — Qwen / Qwen 1.5 / Qwen 2 / Qwen 2.5 chat models alongside the QwQ-32B-Preview (November 2024) and QwQ-32B (March 2025) reasoning experiments that proved the recipe later folded into Qwen3.

Model
QwQ-32B
Qwen/QwQ-32B
Reasoning
Available
Mar 5, 2025
Production reasoning model. 32B dense, Apache 2.0. Reported parity with DeepSeek-R1-671B at 20× fewer parameters on math + coding.
  • Released March 5, 2025; the announcement is at qwenlm.github.io/blog/qwq-32b; HuggingFace card: Qwen/QwQ-32B.
  • 32B dense reasoning model, Apache 2.0. Per Alibaba's testing, reportedly competitive with DeepSeek-R1 (671B total / 37B active) on AIME / MATH benchmarks at a 20× reduction in parameters.
  • The production successor to QwQ-32B-Preview; trained with reinforcement learning to incentivize chain-of-thought reasoning, parallel to the DeepSeek-R1 / Magistral / o1 recipe.
  • The last standalone Qwen reasoning model before the Qwen3 hybrid-mode convergence in April 2025; on this page's convergence reading, no further Reasoning-family rows are expected unless Alibaba revives the standalone QwQ track.
Model
Qwen2.5-Omni-7B
Qwen/Qwen2.5-Omni-7B
Specialized
Available
Mar 2025
End-to-end multimodal: text + image + audio + video in one architecture. 7B Apache 2.0. The first natively-omni Qwen.
  • Released March 2025; the announcement is at Alibaba Cloud Community.
  • First end-to-end multimodal Qwen — processes text, images, audio, and video in a single 7B architecture, with streaming responses across modalities.
  • Apache 2.0 license; HuggingFace at Qwen/Qwen2.5-Omni-7B.
  • Positioned as Alibaba's open answer to GPT-4o's omni-modality release; the recipe was generalized into Qwen3.5's native multimodality at the architecture level. The Omni product slot itself passed to Qwen3-Omni (September 22, 2025) and then to Qwen3.5-Omni (March 30, 2026).
Model
Qwen2.5-VL family
Qwen/Qwen2.5-VL-{3B, 7B, 32B, 72B}-Instruct
Specialized
Available
Jan 2025
Vision-language models built on Qwen2.5. Four sizes (3B / 7B / 32B / 72B). Document parsing, video understanding. Apache 2.0.
  • Released January 2025; HuggingFace cards on the Qwen org.
  • Four sizes (3B, 7B, 32B, 72B) built on the Qwen2.5 base; designed for image / document / chart / video understanding.
  • Strong document-parsing and table-recognition results; 32B and 72B variants positioned as production-grade vision-language flagships.
  • License: Apache 2.0 for the 7B and 32B, with the 72B following the same Qwen License pattern as Qwen2.5-72B and the 3B carrying the narrower research-only Qwen Research License.
  • Superseded in the vision-language slot by Qwen3-VL (September 23, 2025); the weights are still served.
Model
Qwen2.5-Max
qwen-max-2025-01-25 — closed-weights, DashScope only
Flagship
Legacy
Jan 29, 2025
First proprietary Qwen-Max. Large-scale MoE on 20T tokens. Pitched as a DeepSeek-V3 competitor at launch.
  • Released January 29, 2025 — two days after the DeepSeek-R1 Nvidia-stock-crash episode; the announcement is at qwenlm.github.io/blog/qwen2.5-max; coverage in SiliconANGLE and VentureBeat.
  • Large-scale Mixture-of-Experts pretrained on 20 trillion tokens; further post-trained with curated SFT and RLHF.
  • Proprietary, closed-weights, DashScope-only. Pitched at launch as competitive with DeepSeek-V3, Llama-3.1-405B, and the Qwen2.5-72B open-weights flagship.
  • The first Qwen-Max release; established the proprietary-flagship product slot that Qwen3-Max and Qwen 3.6-Max-Preview would later occupy.
Model
QwQ-32B-Preview
Qwen/QwQ-32B-Preview
Reasoning
Legacy
Nov 28, 2024
First Qwen reasoning model. 32B Apache 2.0. Beat o1-preview on AIME and MATH per Alibaba's testing.
  • Released November 28, 2024; coverage in InfoQ and TechCrunch; HuggingFace card: Qwen/QwQ-32B-Preview.
  • 32B-parameter reasoning model under Apache 2.0; the first Qwen release explicitly positioned against OpenAI's o1 series as an open-weights reasoning alternative.
  • Per Alibaba, beat o1-preview on AIME and MATH benchmarks at launch, demonstrating that 32B parameters could match much larger reasoning models on math / coding tasks.
  • Superseded by the production QwQ-32B (March 2025); status is Legacy.
Model
Qwen2.5-Coder family
Qwen/Qwen2.5-Coder-{0.5B, 1.5B, 3B, 7B, 14B, 32B}-Instruct
Specialized
Available
Nov 12, 2024
Six coding-specialist sizes (0.5B–32B). 32B-Instruct reportedly competitive with GPT-4o on HumanEval. Apache 2.0.
  • Released November 12, 2024 with six sizes (0.5B, 1.5B, 3B, 7B, 14B, 32B). HuggingFace cards on the Qwen org.
  • The flagship Qwen2.5-Coder-32B-Instruct was reportedly competitive with GPT-4o on HumanEval / MBPP at launch — the strongest open-weights coding model of late 2024 alongside DeepSeek-Coder-V2.
  • License: Apache 2.0 for the 0.5B / 1.5B / 7B / 14B / 32B variants — including the flagship 32B — with only the 3B carrying the narrower research-only Qwen Research License.
  • Superseded as the recommended open-weights coding model by Qwen3-Coder (July 2025).
Model
Qwen2.5 family (the “Party”)
Qwen/Qwen2.5-{0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B}, base + instruct
Flagship
Legacy
Sep 19, 2024
“A Party of Foundation Models.” Seven dense sizes. 18T tokens. Most variants Apache 2.0; the 72B under the Qwen License and the 3B under the stricter research-only terms.
  • Released September 19, 2024 as “A Party of Foundation Models”; the announcement is at qwenlm.github.io/blog/qwen2.5; technical report at arXiv 2412.15115.
  • Seven dense sizes (0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B) released together — the broadest single-day Qwen release wave.
  • 18 trillion pretraining tokens, ~2.5× Qwen2's corpus, with a focus on knowledge, coding, and mathematics.
  • License: Apache 2.0 for the 0.5B / 1.5B / 7B / 14B / 32B variants; the bespoke Qwen License for the 72B, and the narrower research-only Qwen Research License for the 3B.
  • Established Qwen2.5 as the dominant open-weights mid-tier line through late 2024 and most of 2025; superseded as the flagship by Qwen3 in April 2025.
Model
Qwen2-VL
Qwen/Qwen2-VL-{2B, 7B, 72B}-Instruct
Specialized
Legacy
Aug 29, 2024
First Qwen2-era vision-language. Three sizes (2B / 7B / 72B). Variable-resolution visual encoder.
  • Released August 29, 2024; HuggingFace cards on the Qwen org.
  • Three sizes (2B, 7B, 72B) with a dynamic-resolution visual encoder for variable-aspect-ratio image inputs and longer-video understanding.
  • License: Apache 2.0 at 2B and 7B; the 72B is not. Qwen/Qwen2-VL-72B-Instruct ships a Qwen LICENSE AGREEMENT — commercial use with restrictions — the same top-of-range carve-out Qwen2 and Qwen2-Math make (checked September 4, 2026).
  • Built on the Qwen2 base; superseded by Qwen2.5-VL (January 2025) and Qwen2.5-Omni (March 2025).
Model
Qwen2-Audio + Qwen2-Math
Qwen/Qwen2-Audio-7B-Instruct, Qwen/Qwen2-Math-{1.5B, 7B, 72B}-Instruct
Specialized
Legacy
Aug 8, 2024
First Qwen2-era audio (speech understanding) and math specialists. Apache 2.0 — except Qwen2-Math-72B, which keeps the Tongyi-Qianwen License, the same carve-out the Qwen2 flagship makes at 72B.
  • Released August 8, 2024 as a paired specialized launch.
  • Qwen2-Audio-7B-Instruct — first Qwen audio-understanding model, supporting speech transcription, voice chat, and audio-content QA.
  • Qwen2-Math — three sizes (1.5B, 7B, 72B) targeting STEM-reasoning workloads. The recipe later folded into general-purpose Qwen3 reasoning capability.
  • License: Apache 2.0 at every size but one. Qwen2-Audio-7B-Instruct and Qwen2-Math at 1.5B and 7B are plain Apache 2.0, but Qwen/Qwen2-Math-72B-Instruct and its base checkpoint carry license_name: tongyi-qianwen, and the repo's LICENSE is the Tongyi Qianwen License Agreement — commercial use with restrictions. That is the same 72B carve-out Qwen2 itself made, applied to the specialized line; it is easy to miss because the release was announced as an open-source pair (checked September 4, 2026).
  • Superseded by general-purpose Qwen3 / Qwen3.5 instruction tuning by 2025.
Model
Qwen2 family
Qwen/Qwen2-{0.5B, 1.5B, 7B, 57B-A14B, 72B}, base + instruct
Flagship
Legacy
Jun 6, 2024
Five sizes including the first Qwen MoE (57B-A14B). 27 additional languages. 128K context. Apache 2.0 except 72B.
  • Released June 6, 2024; the announcement is at qwenlm.github.io/blog/qwen2; technical report at arXiv 2407.10671.
  • Five sizes — four dense (0.5B, 1.5B, 7B, 72B) and the first production Qwen MoE (Qwen2-57B-A14B, 57B total / 14B active). 128K context across the family.
  • Pretrained on data covering 27 additional languages beyond English and Chinese; Qwen2-7B-Instruct and Qwen2-72B-Instruct extended context to 128K tokens.
  • License turn: Qwen2-0.5B / 1.5B / 7B / 57B-A14B shipped under Apache 2.0, the first Qwen flagship release with broad permissive coverage. Only Qwen2-72B retained the original Tongyi-Qianwen License.
  • The architectural foundation for everything that followed in the Qwen2.x line.

The Apache 2.0 turn — June 2024. Above this line: most of the Qwen line ships under Apache 2.0, with only the largest size in each generation typically retaining the bespoke Tongyi-Qianwen License. Below: the founding lineage — Qwen 1 (August–November 2023) and Qwen 1.5 (February–March 2024) — entirely under the Tongyi-Qianwen License, the bespoke Alibaba license that allowed academic and commercial use with restrictions. The Qwen2 launch was the structural commitment to permissive open-source that made Qwen the dominant HuggingFace-leaderboard line through 2025.

Model
Qwen 1.5 family + Qwen1.5-MoE
Qwen/Qwen1.5-{0.5B, 1.8B, 4B, 7B, 14B, 32B, 72B, 110B, MoE-A2.7B}-Chat
Flagship
Legacy
Feb 2024
Qwen 1.5 and Qwen 1.5-MoE-A2.7B (March 28, 2024). First Qwen MoE. Tongyi-Qianwen License — but the 0.5B / 1.8B / 4B checkpoints are research-only, base and chat alike.
  • Qwen 1.5 family released February 2024 across multiple sizes (0.5B, 1.8B, 4B, 7B, 14B, 32B, 72B, 110B); extended context support and improved instruction-following over Qwen 1.
  • Qwen1.5-MoE-A2.7B followed March 28, 2024 as the first Qwen Mixture-of-Experts release — ~14B total parameters with 2.7B active per token.
  • License: two licenses, and the split does not run by size. The 7B / 14B / 32B / 72B / 110B / MoE-A2.7B checkpoints carry the bespoke Tongyi-Qianwen License — commercial use with restrictions, not OSI-compliant, the same terms as Qwen 1. The 0.5B, 1.8B, and 4B checkpoints instead carry the Tongyi Qianwen RESEARCH License, which grants no commercial rights at all. So “the smaller sizes are more permissive” is exactly backwards here (all nine sizes, base and chat, checked September 4, 2026).
  • One caveat on the 32B, where Alibaba's own metadata contradicts itself. Qwen1.5-32B (base) is tagged license_name: tongyi-qianwen-research on its HuggingFace card, which reads as research-only. The LICENSE file actually shipped in that repo is not the research agreement: it is the ordinary Tongyi Qianwen License Agreement, byte-for-byte identical to the one in Qwen1.5-32B-Chat, with no occurrence of the word “research” anywhere in it. Every checkpoint that is genuinely research-only — the 0.5B, 1.8B and 4B, base and chat — carries a LICENSE that opens “Tongyi Qianwen RESEARCH LICENSE AGREEMENT” and grants rights “FOR NON-COMMERCIAL PURPOSES ONLY.” The 32B base carries neither. I read that as a card-tag error rather than a real license split, but the two surfaces disagree and only Alibaba can settle it, so check both before relying on the 32B base commercially (checked September 4, 2026).
  • Superseded as the recommended open-weights line by Qwen2 (June 2024) and the broader Apache 2.0 turn that followed.
Model
Qwen 1 family (1.8B / 7B / 14B / 72B)
Qwen/Qwen-{1_8B, 7B, 14B, 72B}, -Chat variants
Flagship
Legacy
Aug 3, 2023
The lab's debut. Qwen-7B (Aug 3), Qwen-14B (Sep 25), Qwen-1.8B + Qwen-72B (Nov 30). Tongyi-Qianwen License — except the 1.8B, which is research-only.
  • Qwen-7B released August 3, 2023 on ModelScope and HuggingFace; the debut of the line. Qwen-14B followed September 25, 2023; Qwen-1.8B and Qwen-72B on November 30, 2023.
  • Four sizes total: 1.8B, 7B, 14B, 72B, each with base (`-Base`) and chat (`-Chat`) variants. GitHub: QwenLM/Qwen.
  • Followed Alibaba Cloud's April 2023 Tongyi Qianwen launch — the lab's response to ChatGPT, formally announced four months before the first open-weights release.
  • License: bespoke Tongyi-Qianwen License for 7B, 14B, and 72B — commercial use permitted with restrictions, not OSI-compliant. The 1.8B is the exception: Qwen/Qwen-1_8B ships under the separate Tongyi Qianwen RESEARCH License (release date November 30, 2023), which carries no commercial grant at all — so the smallest size in the debut wave is the most restricted, not the most permissive (checked September 2, 2026). Established the licensing pattern that later partially shifted to Apache 2.0 at Qwen2.
  • The architectural foundation for everything that followed in the Qwen line through Qwen3.6.

Click any row to expand. Each row has a stable id for sharing — e.g. /ai/qwen/versions/#qwen-agentworld, #qwen-3-6-27b, #qwen-3-next, #qwen-3-vl, #qwen-3-coder-next, #qwen-3, #qwen-2-5, #qwq-32b. Qwen blog: qwen.ai/research (the older qwenlm.github.io/blog is retired, frozen at September 2025, and kept here only for the pre-October-2025 permalinks); HuggingFace org: huggingface.co/Qwen; GitHub org: github.com/QwenLM; release log and model strings: docs.qwencloud.com/changelog/models.

The April 2023 Tongyi Qianwen launch

Alibaba Cloud formally launched Tongyi Qianwen (通义千问, “truth from a thousand questions”) in April 2023 as the company's response to ChatGPT, two months after Baidu's Ernie Bot launched and roughly five months after OpenAI's November 2022 ChatGPT release. The model was first demonstrated by then–Alibaba CEO Daniel Zhang at the Alibaba Cloud Summit on April 11, 2023 and rolled out to enterprise customers through the Tongyi product family on Alibaba Cloud.

The first open-weights release — Qwen-7B — followed on August 3, 2023. The four-month gap between the Tongyi consumer-product launch and the first open-weights release was characteristic of Alibaba's strategy: ship the proprietary chatbot first to enterprise customers via Alibaba Cloud, then open-source the underlying model line to developer communities for ecosystem effects. The same hybrid pattern ran unbroken until August 2026: open-weights flagships on HuggingFace alongside a closed Qwen-Max line on Alibaba's hosted API. That ended on August 12, 2026, when the Max line shipped weights for the first time — on its own license, and without the vision input the hosted model keeps.

The Apache 2.0 turn — from Tongyi-Qianwen License to permissive open-source

Qwen's licensing has moved through three eras — a bespoke house license, a broad Apache 2.0 turn, and a 2026 return to bespoke terms at the frontier — and it has never been as tidy inside an era as the era labels suggest. The Qwen 1 lineage (Qwen-7B / 14B / 72B, August–November 2023) shipped under the bespoke Tongyi-Qianwen License — an Alibaba-authored license with permissive terms for academic use and commercial use with restrictions, but not OSI-compliant. The Qwen 1.5 family in February 2024 continued the same pattern at 7B, 14B, 72B, 110B, and MoE-A2.7B. What is easy to miss, because it inverts the intuition that smaller weights come with fewer strings, is that the small checkpoints in both waves are the restricted ones: Qwen-1.8B, and Qwen1.5 at 0.5B, 1.8B, and 4B, ship under a separate Tongyi Qianwen RESEARCH License that grants no commercial rights at all. There is no size rule to lean on here; the per-checkpoint license is the answer (verified across all of them September 4, 2026). And the card is not always that answer: Qwen1.5-32B's base checkpoint is tagged research-only on HuggingFace while the LICENSE file in the same repo is the ordinary commercial agreement, byte-identical to its -Chat sibling's. When the tag and the license text disagree, read the license text.

The licensing turn arrived with Qwen2 on June 6, 2024. Most of the Qwen2 sub-family (0.5B, 1.5B, 7B, 57B-A14B) shipped under Apache 2.0 — the first Qwen flagship release with broad permissive coverage. Only Qwen2-72B retained the Tongyi-Qianwen License. The pattern continued through Qwen2.5 (September 19, 2024), where the 0.5B / 1.5B / 7B / 14B / 32B variants shipped Apache 2.0 while the 72B retained the commercial-with-restrictions Qwen License and the 3B — alone in the middle of an otherwise permissive range — took the stricter Qwen Research terms, with no commercial grant. The same 3B carve-out repeats across Qwen2.5-VL, Qwen2.5-Coder, and Qwen2.5-Omni.

Qwen3 (April 28, 2025) was the structural commitment: the entire Qwen3 family — six dense sizes from 0.6B to 32B, plus the 30B-A3B and 235B-A22B MoE variants — shipped Apache 2.0. Every open-weights Qwen flagship release since (Qwen3-Coder, Qwen3.5, Qwen3.5-Plus's open-weights variants, Qwen3.6-27B, Qwen3.8-27B, Qwen-Image-2512) has shipped Apache 2.0. The proprietary Qwen-Max-line (Qwen2.5-Max and Qwen3-Max through Qwen3.7-Max / Qwen3.7-Plus, with their hosted Plus / Flash siblings) runs as a parallel commercial track on DashScope, but the open-weights story has been Apache-2.0-or-permissive across every release since Qwen2. In July 2026 the specialized track picked up a hosted-only wrinkle: Qwen-Audio-3.0-Realtime (July 15, 2026), Qwen-Audio-3.0-TTS (July 20, 2026), Qwen-Image-3.0 (July 21, 2026), and Qwen-Audio-3.0-ASR-Flash (July 30, 2026) all shipped as API- or chat-only services with no license text and no weights — the first specialized lines to debut closed while their open-weights predecessors (Qwen3-TTS, Qwen3-Omni, Qwen-Image-2.0, Qwen3-ASR) stayed Apache 2.0.

In August 2026 the flagship track moved the other way, and added two more license conventions doing it. Announcing Qwen3.8-Max on August 3, 2026, Alibaba wrote that the launch “marks the first time we will open-source the weights of a Qwen-Max-class model” (Qwen3.8-Max: A New Bar for Coding and Cowork). The weights arrived on August 12, 2026 as Qwen/Qwen3.8-2.4T-A95B, ending six generations of an entirely closed Max line running from Qwen2.5-Max in January 2025 — but not under Apache 2.0. The repo carries a bespoke Qwen3.8-Max License: broad rights to use, modify, host, fine-tune, and sell, conditioned on prominent model-name attribution for products above 100 million MAU or $20 million monthly revenue, and on obtaining a separate commercial license if the licensee runs a model-as-a-service or AI-work-assistant business above $50 million of trailing-twelve-month group revenue. The checkpoint is also text-only; vision input stays behind the hosted API.

Two days later, on August 14, 2026, Qwen3.8-27B shipped under plain Apache 2.0 — so the split now runs by size rather than by track: the deployable sizes stay permissive, the frontier-scale checkpoint carries revenue-share terms. HuggingFace's State of Open Models: Summer 2026 (August 14, 2026) reads it as a sector-wide turn rather than an Alibaba quirk: of 178 Chinese releases above 20B parameters in 2026, 59% are Apache 2.0 and 22% MIT with almost none carrying non-commercial restrictions, “however, in the last few weeks, we started to see a change on this trend for the really large models, with Kimi K3 and Qwen 3.8 2.4T starting to include some non-commercial restrictions and revenue share requirements to their licenses.”

Twelve days after that report, the size-band reading stopped holding. Qwen3.8-Flash-Next (August 26, 2026) is a 125-billion-parameter model activating 6 billion — a size squarely in the band HuggingFace had just described as staying permissive, and one people actually self-host — and it shipped under a fifth convention, the Qwen Community License 1.0. The text is a near-copy of the Qwen3.8-Max License, with one substantive change that runs the stricter way: where the Max license triggers its separate-commercial-license requirement only once a model-as-a-service or AI-work-assistant licensee clears $50 million of trailing-twelve-month group revenue, the Community license drops the revenue threshold entirely and applies to any such business at any size. The 100-million-MAU / $20-million-monthly-revenue attribution condition and the internal-use carve-out are identical. So the line running through Alibaba's August is not really about parameter count — it is that Qwen now reserves the resale-inference and coding-assistant markets on its newest architectures, at every scale it has released them, while the established generations (Qwen3.8-27B, Qwen3.6, Qwen3.5, Qwen3) stay plain Apache 2.0. Two days after that, Qwen-Drive-1.0-4B (August 28, 2026) shipped under plain Apache 2.0 — and it is built on Qwen3.5-4B, which fits the reading rather than breaking it: the bespoke terms so far attach to the newest architectures, not to derivative work on the settled ones.

The Qwen3 hybrid-reasoning era — April 28, 2025

Qwen3 launched on April 28, 2025 as Alibaba's first frontier-AI line with hybrid reasoning architecture. The release shipped eight models simultaneously — six dense (0.6B, 1.7B, 4B, 8B, 14B, 32B) and two MoE (30B-A3B and the flagship 235B-A22B) — all under Apache 2.0, all sharing a single architecture that supports both thinking-mode chain-of-thought reasoning and non-thinking-mode fast responses. The announcement is at qwenlm.github.io/blog/qwen3; coverage in TechCrunch and Alibaba Cloud Community.

The Qwen3 architecture absorbed the standalone QwQ reasoning track into the Flagship V-series. The QwQ line had run for five months — QwQ-32B-Preview in November 2024, the production QwQ-32B in March 2025 — as Alibaba's open-weights answer to OpenAI's o1 series and DeepSeek's R1. Qwen3's hybrid-mode architecture made the standalone Reasoning family redundant; no further QwQ releases have shipped in the year since, and I don't expect to add another Reasoning row here unless Alibaba revives the standalone track.

Qwen3 was trained on 36 trillion tokens — double Qwen2.5's pretraining corpus — with native multilingual support across 119 languages and dialects, the broadest language coverage of any frontier-AI line at the time. The Qwen3.5 family released ten months later (February 2026) extended that to 201 languages and added native multimodality across text + image + video; Qwen3.6-27B (April 2026) added the Gated DeltaNet hybrid architecture and Thinking Preservation. The Qwen3 architecture and its descendants remain the load-bearing recipe for everything Alibaba has shipped since.

The HuggingFace-leaderboard dominance

Across late 2024 and most of 2025, Qwen-derived models held a majority of the top-trending positions on the HuggingFace Open LLM Leaderboard and the various community-derived benchmark trackers. The pattern was driven by a combination of three factors: the Apache 2.0 license (which let community fine-tuners commercially redistribute their derivatives), the breadth of base sizes (0.5B through 235B in the Qwen3 family alone), and the strength of Qwen2.5-Coder / QwQ / Qwen3 on math + coding benchmarks specifically (which weighted heavily in the leaderboard rankings). At various points in 2025, more than half of the top 20 trending HuggingFace models were Qwen-derived. The dominance shifted slightly in 2026 as DeepSeek-V4 / Llama 5-equivalents / Mistral 3 absorbed share, but Qwen-derivatives have remained a large fraction of the top-trending HuggingFace models through August 2026.

HuggingFace put numbers on it in State of Open Models: Summer 2026 Observations (August 14, 2026): Qwen-based models account for 151,448 derivatives on the Hub — 2.6× Meta's total footprint and 4.7× the Llama repositories specifically, with Google second at 82,506 — growing at roughly 180–210 new repositories a day across the first seven months of 2026. HuggingFace credits three things: a steady release cadence, coverage across sizes, and Apache 2.0. The third-largest derivative source is Unsloth, a community account publishing quantized builds, most of them Qwen. Alibaba's own reading of the same report (August 17, 2026) claims 460-plus open-sourced models, 300,000-plus derivatives and 3 billion cumulative downloads — a broader count than HuggingFace's Hub-only figures, and one the vendor is doing itself.

The U.S. chip export-control context

Qwen training has been constrained by the same U.S. chip export-control regime that shapes DeepSeek's training environment. The October 7, 2022 Department of Commerce export controls restricted top-tier Nvidia AI-GPU exports to China; Alibaba Cloud was already on various U.S. Entity List adjacencies prior to that, with subsequent expansions through 2023–2025 tightening the procurement environment. Like DeepSeek, Alibaba Cloud built Qwen's training infrastructure on a mix of Nvidia H800 chips procured during the gap before the October 2023 H800 ban, and on Chinese domestic alternatives (Huawei Ascend, Cambricon).

The empirical record — trillion-parameter Qwen3-Max trained on 36T tokens, Qwen3.6-27B with state-of-the-art coding benchmarks — demonstrates that the Qwen team has continued training at frontier scale despite the controls, and like DeepSeek has been the subject of Department of Commerce inquiries about the chip procurement that supported specific runs. Coverage in CSIS and South China Morning Post covers the broader policy environment.

At the May 2026 Alibaba Cloud Summit — the same event that launched Qwen3.7-Max — Alibaba pushed further toward domestic alternatives, unveiling the Zhenwu M890, a custom AI accelerator from its T-Head semiconductor subsidiary (144 GB on-chip memory, 800 GB/s interchip bandwidth, ~3× the performance of the prior Zhenwu 810E) purpose-built for agent workloads, alongside the Panjiu AL128 server packing 128 accelerators per rack. Alibaba outlined a multi-year in-house silicon roadmap (a V900 successor in Q3 2027 and a J900 in Q3 2028) and said T-Head has shipped more than 560,000 Zhenwu units to 400+ customers to date. The company has pledged more than 380 billion yuan (~$53 billion) across cloud and AI infrastructure over three years. Coverage in South China Morning Post, CNBC, and Reuters / Business Standard (May 20, 2026).

On June 8, 2026 the U.S. Department of Defense published an updated Section 1260H list of "Chinese military companies," adding 65 entities — 17 new parent-level listings and 48 subsidiaries — and naming Alibaba for the first time, alongside Baidu, BYD, NIO, Unitree, and TP-Link. The DoD described Alibaba and Baidu as contributors to the Chinese defense industrial base. The 1260H list is distinct from the Commerce Department's Entity List and imposes no export licensing requirement of its own; its effect is that the Defense Department is barred from contracting directly with listed companies, and from procuring their products or services through third parties beginning in June 2027. Alibaba disputes the designation and has taken it to court: on June 23, 2026 it sued the Defense Department in the U.S. District Court for the Northern District of California (Alibaba Group Holding Ltd. v. U.S. Department of Defense, No. 5:26-cv-06227), saying it does not work with the Chinese military and asking to be removed from the list. A companion provision — barring the Pentagon from working with any company whose lobbyists also represent a 1260H entity — took effect at the end of June and led all of Alibaba's registered lobbyists to withdraw; on July 5, 2026 Judge Eumi K. Lee ordered the department not to apply that lobbying restriction to Alibaba until she rules on the company's motion or 60 days after a hearing on it, whichever comes first. That order is temporary and does not touch the listing itself, which stood at this writing. Coverage in CNBC (June 9, 2026), a client alert from WilmerHale (June 11, 2026), and Bloomberg via Fortune (July 5, 2026); the docket is tracked at the Civil Rights Litigation Clearinghouse. At WAIC 2026 (July 20, 2026) Alibaba's T-Head unit open-sourced its SAIL AI software stack, optimized for the Zhenwu accelerators, continuing the domestic-silicon push.

The export-control story gained a China-side counterpart in July 2026. Per Financial Times reporting on July 20, 2026 (picked up by Reuters; see also Tom's Hardware), China's Ministry of Commerce has consulted leading AI and chip firms — Alibaba, ByteDance, and Z.ai among them — on possible export controls covering their most advanced AI models, including whether foreign users should be allowed to download model weights and whether key training data may be transferred overseas, alongside limits on foreign manufacture of advanced chips designed by Huawei, Alibaba, and ByteDance. The measures under discussion would ride the next revision of China's catalogue of technologies restricted from export; nothing has been decided, and regulators were still gathering industry feedback at this writing. A restriction on foreign downloads of open-weights models would apply directly to the Apache 2.0 Qwen releases this page tracks — I'll record whatever rule actually lands.

Where to run Qwen

Qwen is among the most widely-deployed AI lines because the open-weights releases are Apache 2.0 across nearly every size and the proprietary releases are available through Alibaba Cloud's Model Studio with OpenAI- and Anthropic-compatible APIs. Inference paths through 2025–2026 break into four categories.

Alibaba Cloud first-party. Qwen Chat is the consumer chat surface. Model Studio (formerly DashScope) is the long-standing developer API endpoint, OpenAI-API-compatible and serving both the open-weights and the proprietary Max-line models. The proprietary Max-line is exclusive to this surface. Since May 26, 2026 it has a front end built for agents: Alibaba Cloud launched Qwen Cloud for international markets in Singapore, an AI-native platform aggregating 150-plus model APIs — Qwen plus third-party lines including DeepSeek, GLM, Kimi, Wan, and HappyHorse — behind one API key, with three entry points (agent-readable Skills, a CLI, and a website) and a per-model marketplace listing price, context, and rate limits (Alibaba Cloud Community, May 28, 2026). Its model-releases changelog is now the fastest-updating first-party record of what has shipped. The older Model Studio docs lagged it badly through the summer, and unevenly: the recommended-models page sat frozen at a July 15, 2026 stamp for four weeks after Qwen3.8-Max went generally available, and the billing page stayed frozen at that stamp for six. Both have since caught up, and by September 2, 2026 both carried a same-day stamp along with the qwen3.8-max and qwen3.8-flash strings — so all three first-party surfaces are current again. The episode is a reminder that a frozen doc page answers “no” to every question you ask it, so the stamp is worth reading before the content. The underlying HTTP endpoint is unchanged — dashscope-intl.aliyuncs.com/compatible-mode/v1. A third first-party surface arrived on August 3, 2026: QwenWork, a workplace agent platform that entered public beta in China offering Qwen3.8-Max as its Flagship model tier and slated for embedding in DingTalk; an international edition followed on August 26, 2026. One consumer surface sits outside Alibaba's own apps: on July 15, 2026 the Cyberspace Administration of China approved Apple Intelligence for launch in China with Qwen as its system-level language engine across iOS / iPadOS / macOS / visionOS (TechCrunch, July 16, 2026).

Self-host from HuggingFace. Download from the Qwen org and run with vLLM, SGLang, llama.cpp, or Ollama. The Apache 2.0 open-weights flagships (Qwen3.8-27B, Qwen3.6-27B, Qwen3.5 family, Qwen3 family, QwQ-32B) self-host without commercial restriction. Three smaller sets carry conditions: the Qwen License variants (Qwen2-72B, Qwen2.5-72B, Qwen2.5-3B) require attestation of the bespoke terms; Qwen3.8-2.4T-A95B — the open Max checkpoint, and at 2.4 trillion parameters not a self-host most people will attempt anyway — requires a separate commercial license above $50 million of trailing-twelve-month revenue if you are reselling inference; and Qwen3.8-Flash-Next, which at 125B total and 6B active is a realistic self-host, requires that separate license for any model-as-a-service or AI-work-assistant business with no revenue floor at all. Running it for your own internal use is explicitly carved out of both.

Hyperscalers. AWS Bedrock and Azure AI Foundry have added Qwen SKUs across 2025–2026; ModelScope (Alibaba's own model-hub) hosts the broadest set. NVIDIA NIM has Qwen variants for the most-served sizes.

Hosted-inference providers. Together AI, Fireworks, OpenRouter, SiliconFlow, Groq. Most providers serve the Apache 2.0 lineage with similar latency / cost characteristics; the Tongyi-Qianwen-License variants (mostly the 72B and 3B sizes) are typically not carried by Western inference providers due to license-attestation overhead.

People who shaped Qwen

The Qwen / Tongyi Lab team is structured inside Alibaba Cloud rather than as a standalone lab. Junyang Lin (Lin Junyang) ran the Qwen project from the Tongyi Lab's formation in late 2022 — through the April 2023 Tongyi Qianwen launch and every release up to the Qwen3.5 small-model wave — and then left abruptly, posting “me stepping down. bye my beloved qwen.” on X in the early hours of March 4, 2026, with March 7 as his last day. Several Qwen engineers went in the same window, including post-training lead Yu Bowen. Alibaba Cloud CTO Jingren Zhou (Zhou Jingren), who built the Tongyi Lab and set the open-source strategy, took over oversight of the team, and Zhou Hao — previously a senior staff research scientist at Google DeepMind — joined as head of post-training research. Reporting attributed the split to disagreements over lab structure (Lin favored keeping pretraining and post-training integrated, against a reorganization separating them) and over the Qwen team's control of its own AI infrastructure: South China Morning Post (March 4, 2026) and Caixin Global (March 10, 2026). The release cadence did not slow: Qwen3.6-35B-A3B, Qwen3.6-27B, the Qwen3.7 generation, Qwen-AgentWorld, and the July 2026 wave all shipped after the departure.

Eddie Wu — CEO of Alibaba Group since September 2023; has publicly framed AI as Alibaba's strategic priority above e-commerce / cloud / logistics, with multi-year capex commitments to Qwen training infrastructure. He took a more direct hand after Lin's exit: a foundation-model task force under Wu stood up on March 5, 2026, and on March 16 Alibaba folded its AI teams and products into a new Token Hub Business Group, also led by Wu, sitting alongside Alibaba Cloud and the e-commerce divisions. The reshuffling continued: on June 8, 2026 Alibaba merged the Tongyi Lab with the Future Life Lab (the Taobao-and-Tmall team behind the Happy Horse video model and the Happy Oyster world model) into a new Token Foundry unit, again reporting directly to Wu, while Jingren Zhou moved up to Alibaba group chief scientist — the highest academic title in Alibaba's technology system — and took charge of a newly created AI Future Research Institute focused on frontier research; Zheng Bo, who led the Future Life Lab, brought that team into the merged unit. Sources: South China Morning Post and 36Kr (June 8–9, 2026). No further reshuffle or named-lead change had been reported as of September 8, 2026. Joseph Tsai — Chairman of Alibaba Group; has been the public spokesperson for Alibaba's AI strategy in international forums (the Davos annual meetings, the Bloomberg Tech Summit).

No publicly-named Qwen CTO or founding team in the Western-lab sense. Unlike OpenAI / Anthropic / Mistral / xAI / DeepSeek, Qwen is not structured as a startup-style lab with named founders and a public-facing leadership roster. The team operates under Alibaba Cloud's organizational umbrella, and the per-paper author lists on the Qwen / Qwen2 / Qwen2.5 / Qwen3 technical reports are the closest available roster of named contributors.

The competitive landscape

Qwen is, alongside DeepSeek, one of the two dominant Chinese open-weights AI families through 2024–2026. The closest direct comparators on the open-weights axis are DeepSeek (Chinese, MIT-licensed for the V3 / R1 line and onward, the December 2024 / January 2025 inflection — see DeepSeek Versions), Mistral (French; Apache 2.0 for the Mistral 3 family with a parallel proprietary tier — see Mistral Versions), Meta's Llama (custom Llama Community License, see Llama Versions), and the other Chinese frontier labs (Baidu Ernie, Zhipu GLM, MiniMax, Moonshot Kimi). The closed-weights frontier competitors — ChatGPT, Claude, Gemini, Grok — are the practical benchmark for “is Qwen competitive at frontier scale,” which the Qwen3 / Qwen3-Max / Qwen3.5 / Qwen3.6 release cycle has been answering in the affirmative since April 2025. Qwen's distinguishing variable is the breadth of its specialized track (Coder / VL / Audio / Math / Omni / Image) and the consistency of its Apache 2.0 commitment on the open-weights flagships, both of which continue to underwrite the line's HuggingFace-leaderboard dominance. This page does not attempt a benchmark roundup or a ranking.

Use Qwen

The browser cannot detect which Qwen model you've used — there's no fingerprint or header that exposes it. The block below carries the practical information instead: the current model identifiers, a copy-paste API call, the surfaces where Qwen is available, and the licensing summary.

Current model identifiers

HuggingFace ids on the Qwen org for the open weights; DashScope model strings for the hosted models. Verify against qwencloud.com/models — the model marketplace, which carries the exhaustive current list — and huggingface.co/Qwen.

# Open-weights flagship line (Apache 2.0)
Qwen/Qwen3.8-27B                # newest Apache-2.0 open flagship, native vision-language — Aug 14, 2026
Qwen/Qwen3.6-27B
Qwen/Qwen3.6-35B-A3B            # MoE companion to the dense 27B
Qwen/Qwen3.5-{0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B, 397B-A17B}
Qwen/Qwen3-Next-80B-A3B-Instruct  # ultra-sparse MoE, 3B active of 80B
Qwen/Qwen3-Next-80B-A3B-Thinking

# Open-weights specialized (Apache 2.0)
Qwen/Qwen-Drive-1.0-4B          # autonomous driving: 3D perception + planning
Qwen/Qwen-AgentWorld-35B-A3B    # language world model, 7 agent domains
Qwen/Qwen3-Coder-Next           # coding agent, 80B/3B active
Qwen/Qwen3-Coder-{30B-A3B, 480B-A35B}-Instruct
Qwen/Qwen3-VL-{2B, 4B, 8B, 32B, 30B-A3B, 235B-A22B}-Instruct
Qwen/Qwen3-ASR-{0.6B, 1.7B}    # speech recognition, 52 languages
Qwen/Qwen2.5-Omni-7B           # end-to-end multimodal
Qwen/QwQ-32B                    # reasoning
Qwen/Qwen-Image-2512            # text-to-image

# Open weights under the bespoke Qwen3.8-Max License (not Apache 2.0)
Qwen/Qwen3.8-2.4T-A95B          # open Max checkpoint, TEXT-ONLY — Aug 12, 2026
Qwen/Qwen3.8-2.4T-A95B-FP8      # FP8 build of the same

# Open weights under the Qwen Community License 1.0 (not Apache 2.0)
Qwen/Qwen3.8-Flash-Next         # Qwen4 architecture preview, 125B/6B active — Aug 26, 2026
Qwen/Qwen3.8-Flash-Next-FP8     # FP8 build of the same

# Hosted DashScope model strings (per docs: dots, not dashes)
qwen3.8-max                     # current Max-line flagship (Aug 3, 2026), $2 / $6 per 1M
qwen3.8-2.4t-a95b               # hosted build of the open Max checkpoint, TEXT-ONLY
                                # (Aug 13, 2026), also $2 / $6 per 1M
qwen3.8-flash                   # cost tier (Aug 26, 2026), $0.15 / $0.47 per 1M
qwen3.8-27b                     # hosted Qwen3.8-27B (Aug 19, 2026), $0.50 / $3 per 1M
qwen3.7-plus                    # balanced tier (May 31, 2026)
qwen3.7-flash                   # prior cost tier (Jul 25, 2026), still served
qwen3.7-max                     # prior Max flagship (May 20, 2026), still served
qwen3.6-max-preview
qwen3.6-plus
qwen3.6-flash
qwen3.5-plus
qwen3-max
qwen-max-2025-01-25             # Qwen2.5-Max alias
qwen3-coder-next                # hosted build of the open coding agent, $0.30 / $1.50 per 1M
qwen3.5-omni-plus               # omni-modal (text+image+audio+video), realtime speech — Mar 2026
qwen-image-3.0-pro              # text-to-image, long-text + dense layouts — API since Aug 4, 2026
qwen-image-3.0                  # standard quality/speed variant
qwen-image-2.0-pro              # prior image API — Feb 2026, still served
qwen-audio-3.0-tts-flash        # text-to-speech, ~300ms first packet — Jul 2026
qwen-audio-3.0-tts-plus         # text-to-speech, quality-tuned — Jul 2026
qwen-audio-3.0-realtime-flash   # speech-to-speech, full-duplex WebSocket — Jul 2026
qwen-audio-3.0-realtime-plus    # speech-to-speech, quality-tuned — Jul 2026
qwen-audio-3.0-asr-flash        # speech recognition, clips to 5 min — Jul 30, 2026
qwen-audio-3.0-asr-flash-filetrans   # offline file transcription
qwen-audio-3.0-asr-flash-streaming   # real-time streaming

Quick API call (OpenAI-compatible)

DashScope is OpenAI-API-compatible — point any OpenAI SDK at the same base URL with a Qwen Cloud / Model Studio API key. Replace the placeholder values before running.

$ curl https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions \
    -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model":    "qwen3.8-max",
      "messages": [{ "role": "user", "content": "Hello, Qwen." }]
    }'

Where to run Qwen

Four categories — Alibaba Cloud first-party, self-host from HuggingFace, hyperscalers, and hosted-inference providers. Pricing varies by provider; the open weights are the same across all of them.

# Alibaba Cloud first-party
https://qwen.ai/                            # Qwen Chat consumer chat
https://www.qwencloud.com/                  # Qwen Cloud — model marketplace + API keys
https://docs.qwencloud.com/changelog/models # fastest-updating release log
https://www.alibabacloud.com/help/en/model-studio/   # OpenAI-compatible API

# Self-host from HuggingFace
https://huggingface.co/Qwen                 # every model card lives here
https://github.com/QwenLM                   # GitHub org, technical READMEs
https://github.com/vllm-project/vllm        # production-grade throughput
https://ollama.com/                         # single-binary, easiest entry

# Hyperscalers
AWS Bedrock, Azure AI Foundry, ModelScope, NVIDIA NIM

# Hosted-inference providers
https://www.together.ai/
https://fireworks.ai/
https://openrouter.ai/
https://groq.com/
https://www.siliconflow.com/

Licensing

The line runs several license tiers at once, and they do not track model size — in the Qwen 1 / 1.5 / 2.5 generations it is often the smaller checkpoint that is research-only. Read the LICENSE file in the exact repo before shipping at scale — the HuggingFace card tag is a summary, and on at least one checkpoint it disagrees with the license actually shipped.

# Apache 2.0 — commercial use unrestricted
Qwen-Drive-1.0-4B, Qwen3.8-27B
Qwen-AgentWorld (35B-A3B open weights)
Qwen3.6-27B, Qwen3.6-35B-A3B
Qwen3.5 (every open-weights size), Qwen3 (entire family)
Qwen3-Coder-Next, Qwen3-Next-80B-A3B (Instruct + Thinking)
Qwen3-VL (all sizes), Qwen3-Coder, Qwen-Image-2512
Qwen3-ASR, Qwen3-ForcedAligner, Qwen3-Omni (Instruct + Thinking)
QwQ-32B, QwQ-32B-Preview
Qwen2.5-Omni-7B, Qwen2.5-VL (7B, 32B), Qwen2.5-Coder (all but 3B)
Qwen2.5 (most sizes), Qwen2 (all but 72B)
Qwen2-Audio-7B, Qwen2-Math (1.5B, 7B), Qwen2-VL (2B, 7B)

# Tongyi-Qianwen License — commercial use with restrictions
Qwen2-72B, Qwen2-Math-72B, Qwen2-VL-72B
Qwen2.5-72B, Qwen2.5-VL-72B
Qwen 1 (7B / 14B / 72B)
Qwen 1.5 (7B, 14B, 32B, 72B, 110B, MoE-A2.7B)

# Research-only — no commercial grant at any size
Qwen-1_8B                       # Tongyi Qianwen RESEARCH License
Qwen1.5-{0.5B, 1.8B, 4B}        # base + chat both research-only
                                # (Qwen1.5-32B base is TAGGED research on its
                                #  card, but ships the commercial LICENSE text
                                #  — treated as commercial above.)
Qwen2.5-3B, Qwen2.5-VL-3B,      # the narrower "Qwen Research" variant
Qwen2.5-Coder-3B, Qwen2.5-Omni-3B

# Qwen3.8-Max License — open weights, revenue-share conditions
Qwen3.8-2.4T-A95B (+ -FP8)      # attribution above 100M MAU / $20M mo. revenue;
                                # separate license for MaaS or AI-work-assistant
                                # businesses above $50M trailing-12-mo. revenue

# Qwen Community License 1.0 — open weights, STRICTER on resale
Qwen3.8-Flash-Next (+ -FP8)     # same 100M MAU / $20M mo. attribution condition,
                                # but the separate MaaS / AI-work-assistant license
                                # is required at ANY revenue — no $50M floor.
                                # Internal use is carved out.

# Proprietary — hosted only, no weights for THIS build
Qwen3.8-Max                     # hosted build adds vision input + non-thinking mode;
                                # the open 2.4T-A95B above is the text-only sibling
Qwen3.8-Flash                   # hosted build of Qwen3.8-Flash-Next; adds 1M default
                                # context + built-in tools
Qwen3.7-Max, Qwen3.7-Plus, Qwen3.7-Flash
Qwen-Image-3.0 (Pro + standard) # no weights or license published
Qwen-Audio-3.0-TTS, Qwen-Audio-3.0-Realtime, Qwen-Audio-3.0-ASR-Flash
Qwen 3.6-Max-Preview, Qwen 3.6-Plus, Qwen 3.6-Flash
Qwen3.5-Plus, Qwen3-Max
Qwen2.5-Max

Sources: Qwen blog; github.com/QwenLM; huggingface.co/Qwen; Qwen Cloud model-releases changelog; Qwen Cloud model marketplace; Alibaba Cloud Model Studio docs; research papers on arXiv (Qwen2, Qwen2.5, Qwen3 technical reports); contemporaneous reporting in NYT, FT, Bloomberg, CNBC, South China Morning Post, TechCrunch, VentureBeat, MarkTechPost, SiliconANGLE, Simon Willison. Last updated September 8, 2026.

Mungomash LLC · More AI pages