DeepSeek Versions

High-Flyer, Liang Wenfeng, and the July 2023 founding

DeepSeek was founded on July 17, 2023 in Hangzhou, China, by Liang Wenfeng. The lab's parent and sole funder is High-Flyer, the Chinese quantitative hedge fund Liang co-founded in February 2016. Liang serves as CEO of both companies.

Liang's background is unusual for a frontier-AI founder: he studied electrical engineering at Zhejiang University, began trading equities during the 2008 financial crisis as an undergraduate, and built High-Flyer as an AI-driven quant fund. By 2021, High-Flyer was reportedly using AI exclusively for trading decisions and had become one of the largest quantitative funds in China. The fund's profitability is what underwrote the AI-research investment that produced DeepSeek; reporting in Fortune and ChinaTalk covers the trajectory.

DeepSeek was wholly owned by High-Flyer from incorporation through mid-2026. The April 2026 reports of Tencent and Alibaba investment talks — reshaped over May–June 2026 into a far larger state-anchored round that, per June 17, 2026 reporting, raised more than 50 billion yuan (~$7.4 billion) at a $50 billion-plus valuation (covered in the funding section below) — are the company's first external funding round. The round is structured so Liang retains control through a limited partnership he manages; the High-Flyer relationship stays intact.

Fire-Flyer 2 and the GPU stockpile

Before DeepSeek existed as an independent entity, High-Flyer had been building the GPU infrastructure the lab would inherit. Liang began acquiring Nvidia GPUs at scale starting in 2021, reportedly building a stockpile of around 10,000 Nvidia A100 chips before the U.S. October 7, 2022 export controls first restricted top-tier-AI-GPU exports to China. The pre-controls procurement window is the single most-cited piece of context for why a Chinese lab can train at scale despite the sanctions: the chips were already on the floor before the restriction took effect.

High-Flyer built the Fire-Flyer 2 cluster beginning in 2021 with a reported budget of 1 billion yuan. Per the cluster's published statistics, Fire-Flyer 2 had reached 5,000 PCIe A100 GPUs in 625 nodes with ~96% utilization through 2022, totaling ~56.74 million GPU-hours of capacity used. The cluster was the load-bearing infrastructure for everything DeepSeek shipped from DeepSeek-LLM through V3, and the implicit denominator behind the “$5.576M training cost” figure for V3 — that figure is the marginal compute cost of one training run, not the cumulative R&D and infrastructure cost of building the cluster the run depended on. The dispute over whether the headline cost number is misleading hinges on this distinction.

The V3 / R1 inflection — December 2024 / January 2025

DeepSeek-V3 shipped on December 26, 2024 as a 671B-total / 37B-active MoE model with a disclosed training cost of $5.576 million on 2.788 million H800 GPU hours, pretrained on 14.8 trillion tokens. Performance on broad benchmarks at launch was characterized as competitive with GPT-4o and Claude 3.5 Sonnet. The combination — frontier-adjacent quality, fully open weights, and an order-of-magnitude-lower disclosed compute number — was the data point the AI-infrastructure capex thesis had not previously had to absorb. The technical report is at arXiv 2412.19437.

DeepSeek-R1 followed on January 20, 2025, built on the V3 base, with reasoning capability incentivized through reinforcement learning. The accompanying R1-Zero result — trained with RL only and no supervised fine-tuning — demonstrated emergent chain-of-thought reasoning, self-verification, and reflection behaviors, and was the load-bearing scientific claim of the release. Performance was characterized as comparable to OpenAI's o1 across math, coding, and reasoning benchmarks. The paper is at arXiv 2501.12948; a Nature follow-up published September 17, 2025 is at nature.com/articles/s41586-025-09422-z.

On January 27, 2025, the public-equity reaction to the R1 narrative produced what was at the time the largest single-day loss in U.S. stock-market history: Nvidia fell ~17% and shed approximately $589 billion in market capitalization in a single trading session. The Nasdaq fell ~3% on the day; AI-infrastructure names (Broadcom, Marvell, Vertiv, Constellation Energy) sold off in sympathy. CNBC, Yahoo Finance.

The causation is disputed. Tim Lee at Understanding AI argued that the move was already in motion before the R1 release week, that the disclosed-compute figure excluded prior R&D and the Fire-Flyer build, and that an alternative reading is “efficiency gains expand the inference-compute market faster than they shrink the training-compute market” (the Jevons-paradox response that Microsoft, Meta, and Google all subsequently adopted publicly). DeepSeek did not retract the disclosed-compute figure; the dispute is over the framing, not the number.

The U.S. chip export-control context

DeepSeek's training infrastructure has been scrutinized by U.S. policymakers since the R1 release week. The October 7, 2022 U.S. Department of Commerce export controls restricted top-tier AI-GPU exports to China; Nvidia subsequently produced the H800, a deliberately-degraded H100 variant designed to fall under the export-control thresholds, which it sold legally to Chinese customers including DeepSeek. The H800 was banned in turn in October 2023, but the year-long gap between the original control and the H800 ban was sufficient for DeepSeek to procure the chips it disclosed using to train V3 in 2024. Coverage at CSIS, RAND.

Following the R1 release, the U.S. Department of Commerce opened an inquiry into whether DeepSeek had used U.S. chips not legally exportable to China. House Select Committee chairs Krishnamoorthi and Moolenaar issued a public call to tighten the existing controls in February 2025. Through 2025, several U.S. state governments and federal agencies banned the DeepSeek consumer chatbot on government devices on data-handling grounds (the same regime that had been applied to TikTok); the bans do not apply to the open-weights releases on HuggingFace, which can be self-hosted on Western infrastructure.

The April 2026 V4 release re-opened the chip-controversy docket. Reporting in 2026 alleged that the V4 training run used clusters of Nvidia Blackwell B200 GPUs — a chip class that is comprehensively export-controlled to China — reportedly housed at a data center in Inner Mongolia. In February 2026 a Trump-administration official told Reuters that DeepSeek's latest model was trained on Nvidia's most advanced Blackwell chip in possible violation of the export controls, alleging the company would strip the technical indicators that reveal U.S.-hardware provenance; a senior State Department official separately alleged DeepSeek had supported Chinese military and intelligence operations and used Southeast Asian shell companies to access restricted chips. DeepSeek has not publicly confirmed the chips it used to train V4.

Despite those allegations, DeepSeek has not been added to the Commerce Department's Entity List. Reuters reported on June 17, 2026 that an interagency committee (Commerce, Defense, Energy, State) had approved DeepSeek — along with memory chipmaker CXMT and more than 100 other Chinese firms — for blacklisting last year, but the Trump administration held the listings unpublished to avoid disrupting the fragile U.S.–China trade truce. No entity has been added to the list since October 2025 — a gap CSIS's Philip Luck called the longest between postings in over a decade, and one that still stood as of early September 2026, since every Entity List rule the Bureau of Industry and Security has published since is a revision or a removal. The DeepSeek consumer-chatbot bans on U.S. state-government and federal-agency devices remain in place separately; they do not apply to the open-weights releases on HuggingFace.

On July 7, 2026, Reuters reported, citing three people familiar with the matter, that DeepSeek is developing its own inference chip — silicon for serving trained models rather than training new ones — after roughly a year of talks with chip-design, foundry, and memory partners and private hiring of chip-design engineers. The reported aim is to reduce dependence on both Nvidia, whose most advanced parts are export-controlled to China, and Huawei, whose Ascend silicon reportedly runs DeepSeek's cloud service today. DeepSeek has not publicly confirmed the effort, and no tape-out or foundry partner has been named.

In the meantime the Huawei dependence is deepening rather than shrinking. On September 4, 2026, Bloomberg reported that DeepSeek plans to install at least 160,000 of Huawei's next-generation Ascend 950DT accelerators at the roughly one-gigawatt data center it is building in Inner Mongolia — which would be among the largest known clusters of Huawei AI silicon anywhere, and a concrete step in China's effort to substitute domestic parts for Nvidia's. Per the report the chips are earmarked for running models rather than training them, even though Huawei designed and markets the 950DT for the more demanding training workload as well: DeepSeek has tried training on Huawei parts before and has so far kept Nvidia accelerators for that step. The installation timetable is reported to depend on Huawei's production capacity, and that constraint is sharper than the phrasing suggests — shortages of components such as top-end memory will cap Huawei's 950DT output at the low hundreds of thousands of units this year, filling DeepSeek's order could take more than a year, and DeepSeek has asked Beijing to help press Huawei for a larger and faster allocation. The 160,000 accelerators would cover only one chunk of the site's eventual gigawatt-scale capacity; Bloomberg writes that it is unclear what silicon fills the rest, and that Trump-administration officials have alleged DeepSeek procured Nvidia's top Blackwell parts and installed them in Inner Mongolia — a claim the wire says it has not independently verified. All sources are unnamed, DeepSeek did not respond to requests for comment, and nothing in the report resolves what hardware trained V4.

The DeepSeek License vs. MIT — the licensing turn

DeepSeek's licensing has evolved across two distinct conventions. From DeepSeek-LLM (November 2023) through DeepSeek-V3 (December 2024), the model weights shipped under the bespoke “DeepSeek License” — an OpenRAIL-derived custom license with use-based restrictions (military, surveillance, deceptive content, certain weapons applications) and a separate commercial-license track. The associated GitHub source code repos shipped under MIT separately. This is the same code-vs-weights split Meta uses for the Llama lineage, but DeepSeek's bespoke license is differently shaped and was not OSI-approved. Black Duck's model-license review from January 2025 walks the original terms.

The licensing turn is DeepSeek-R1 (January 20, 2025), which was the first DeepSeek flagship released under the MIT License. DeepSeek-V3-0324 (March 24, 2025) re-released the V3 weights under MIT, retroactively bringing the V-series flagship into MIT-compliance for the post-V3 era. Every subsequent V-series and R-series release — R1-0528, V3.1, V3.1-Terminus, V3.2-Exp, V3.2, V3.2-Speciale, V4-Pro, V4-Flash, the V4-Pro-0813 / V4-Flash-0731 GA builds, and the experimental V4-Flash-Vision-Exp — has shipped under MIT. Janus-Pro (January 27, 2025) also shipped under MIT, as did DeepSeek-OCR (October 2025). DeepSeek-Math-V2 (November 27, 2025) was the first DeepSeek model to ship under Apache 2.0 instead of MIT, followed by its OCR successor DeepSeek-OCR-2 (January 2026).

The pre-R1 specialized models (Coder, Coder-V2, Math, VL, VL2) remain on the original DeepSeek License for the model weights as of this page's publication date; whether DeepSeek will retroactively relicense the older specialized weights to MIT is open. The one release that briefly sat outside both buckets no longer does: DeepSeek-V4-Flash-Vision-Exp was API-only for its first ten days, and its August 31, 2026 weights drop landed under MIT like the rest of the V4 line. For new builds, the practical guidance is “everything from R1 forward is MIT — with DeepSeek-Math-V2 and DeepSeek-OCR-2 the two Apache-2.0 exceptions — while the older specialized models retain the use-restriction terms of the DeepSeek License.” Read the LICENSE-MODEL file in the relevant GitHub repo before shipping at scale.

The 2026 first-external-round funding talks

Through April 2026, DeepSeek had raised no external capital — it was funded entirely by High-Flyer's profits since the July 2023 incorporation. On April 22, 2026, Bloomberg and The Information reported that Tencent and Alibaba were in talks to invest a combined ~$1.8 billion at a $20 billion+ valuation. Tencent had proposed acquiring up to a 20% stake but DeepSeek was reluctant to cede that share of control; Alibaba's role was reportedly smaller. The talks landed two days before the V4 release.

By mid-May 2026 the round had reshaped substantially — from the original $20 billion-plus Tencent / Alibaba framing into a far larger state-anchored round. Early-June 2026 reporting (Reuters, Bloomberg) described a round of roughly $7.4 billion (~50 billion yuan) at a 350–400 billion yuan ($52–59 billion) valuation, with Tencent (~10 billion yuan) and battery maker CATL (~5 billion yuan) named participants and founder Liang Wenfeng contributing roughly 20 billion yuan of his own capital.

By June 17, 2026, the Wall Street Journal and The Information reported that DeepSeek had raised more than 50 billion yuan (~$7.4 billion) in its first external round, valuing the company at more than $50 billion — among the largest AI funding rounds in Chinese history. The deal's defining feature is its control structure: outside investors place their capital into a limited partnership Liang manages rather than buying DeepSeek equity directly, accept a five-year lock-up, and receive no voting rights, so Liang retains control. China's state-backed National Artificial Intelligence Industry Investment Fund is the sole exception — it invested roughly 1 billion yuan directly into the company with voting rights and no lock-up. Reported participants include Tencent, CATL, JD.com, NetEase, Hillhouse, IDG Capital, and Monolith Capital; proceeds are earmarked for compute infrastructure and employee compensation. DeepSeek did not comment publicly. The state-vehicle lead, the founder's large self-contribution, and the voting-rights asymmetry are the load-bearing changes from the original Tencent / Alibaba framing.

Less than a month after that round closed, the cadence accelerated again. On July 14, 2026, the Financial Times reported — picked up by TechCrunch and Bloomberg — that DeepSeek is in talks to raise a further ~$1.5 billion at roughly a $71 billion valuation, and has begun preparing an initial public offering on a mainland Chinese exchange, with a filing targeted as early as the end of 2026 and a debut in 2027. The reported driver is the capital needed to build DeepSeek's own data-center capacity and secure more AI chips. Both the round and the IPO are reported-not-confirmed: DeepSeek has not commented publicly, and no prospectus has been filed.

Reuters put firmer numbers on both items the following week. In wire copy dated July 20, 2026 (carried by The Manila Times), Reuters reported that DeepSeek is planning a fresh round of as much as 50 billion yuan (~$7.4 billion) at a valuation of about 500 billion yuan ($74 billion) — a far larger raise than the ~$1.5 billion the FT had described — and that the June round had closed at a post-money valuation of roughly 450 billion yuan, though filings by two Chinese investors later implied a 350.88 billion yuan (~$52 billion) mark. On the listing, Reuters named Shanghai's Nasdaq-style STAR Market as the venue under early deliberation and reported an internal target to complete an IPO filing before the end of 2026. All sources were unnamed, DeepSeek did not respond to a request for comment, and the wire cautioned that both the terms and the timetable may change.

Five days later the second round stopped. On July 25, 2026, Bloomberg reported — carried by Fortune and Yahoo Finance — that DeepSeek had verbally told prospective investors it was suspending the deal for now, and that they would not be signing the investment agreements they had expected in the days ahead. The reported trigger is not the terms but a leak: a transcript of a meeting Liang held with unidentified parties circulated online, and Chinese outlets including Yicai reported that he had discussed DeepSeek's reliance on Nvidia chips and China's persistent lag behind the U.S. in AI sophistication. Bloomberg did not verify the transcript's authenticity, and DeepSeek did not respond to a request for comment on either the transcript or the fundraising. Bloomberg put the paused round at at least 10 billion yuan (~$1.5 billion) at a pre-money valuation of at least 480 billion yuan (~$71 billion) — the Financial Times' figures rather than the larger ones Reuters described — and reported that negotiations remain fluid and the company may resume the process later. The IPO preparation was described as continuing, with a filing possible before the end of the year.

It resumed twelve days after that. On August 6, 2026, Bloomberg reported that DeepSeek had restarted the round and is now seeking close to $8 billion at a valuation near 500 billion yuan (~$74 billion) — the larger figures Reuters had described in July rather than the ~$1.5 billion the FT reported, and roughly a second raise the size of the first. Monolith Management, an early backer of Moonshot AI, is in talks to join. Reported use of proceeds is DeepSeek's own data-center capacity, led by a large facility in Inner Mongolia. Bloomberg cautioned that the size, timing, and investor list can all still change, and DeepSeek has not commented publicly. Two other moves landed the same day and cut against the lab's cheap-compute reputation: the API price-increase warning noted in the V4 rows above, and a 140.8 million yuan ($20.8 million) investment in humanoid-robot maker Unitree Robotics — 933,399 shares, or 2.31% of the strategic placement in Unitree's Shanghai IPO, under a 36-month lock-up — paired with an agreement to jointly develop AI models for humanoid machines, with each company favoring the other for model-training services and robot purchases respectively (Reuters, from a stock-exchange filing). The partnership is DeepSeek's first substantial move toward embodied AI, a direction its language-first model line has not covered.

Three weeks later the round was reported to be closing. On August 26, 2026, the South China Morning Post reported that DeepSeek was nearing completion of a raise of about 50 billion yuan (~$7.4 billion) at a pre-money valuation near 500 billion yuan (~$74 billion), expected to close before the end of August — which would imply a post-money mark around $81 billion. The Wall Street Journal reported the same $74 billion figure and framed the raise as building a war chest ahead of a listing. Named participants are returning backers Monolith, Shixiang Capital, and CATL; new investors reported to be in talks include CPE, Legend Capital, and Stony Creek Capital, a semiconductor-focused private-equity firm, alongside funds backed by chipmaker GigaDevice and state investment vehicles from Hefei. Reported use of proceeds is roughly a gigawatt of added compute capacity. On the listing, the reporting firms up the venue and the clock: an IPO filing on Shanghai's STAR Market as early as the end of 2026, with a market debut targeted for 2027. CNBC covered the same week how High-Flyer affiliates have taken pre-IPO allocations in CXMT and Unitree, tying Liang's widening balance-sheet activity to sectors Beijing treats as strategic. All sources are unnamed and DeepSeek has not commented. That end-of-August window has now passed without an announced closing — as of September 8, 2026 the most recent reporting still describes the round as nearing completion rather than done, and the August 26 SCMP / WSJ stories remain the newest on it.

The same week put the first hard revenue figures on the table. On August 26, 2026, Reuters — carried by The Standard — relayed a report from The Information that DeepSeek booked roughly 475 million yuan (~$70.7 million) in revenue over the first seven months of 2026, about ten times its full-year 2025 revenue of roughly 47.5 million yuan. The same report puts the seven-month net loss at about 715 million yuan — more than the revenue it sits against — versus a 935 million-yuan net loss for all of 2025, and splits the margins: 82.9% gross margin on the API business against 44.6% company-wide. For scale, the report set those against OpenAI's 39% first-quarter gross margin and Anthropic's, which the outlet said is expected to reach 63% for the year. An earlier report the same summer had put DeepSeek's annualized run-rate at $400–500 million. The Information also reported that DeepSeek hired investment banks to prepare a Shanghai listing for next year immediately after closing the June round — the firmest sourcing yet on the IPO track. DeepSeek did not respond to the outlet's request for comment, and none of these figures are audited or company-published.

Where to run DeepSeek

DeepSeek is widely deployed because the weights are open and the API is OpenAI-compatible. Inference paths through 2025–2026 break into four categories.

DeepSeek's own API. The first-party endpoint at api-docs.deepseek.com is OpenAI-API-compatible, so any OpenAI SDK can be pointed at it with only a base-URL change; it also speaks the Anthropic message format and, since the V4 GA wave, the OpenAI Responses API. Pricing has historically been an order of magnitude cheaper than Western frontier-model APIs (V4-Flash at $0.14 / M input tokens at launch), though the August 16, 2026 move to peak / off-peak billing raised every rate — see the V4 rows above for the current numbers.

Self-host from HuggingFace. Download from the deepseek-ai org and run with vLLM, SGLang, llama.cpp, or Ollama. The full V3 / V3.1 / V3.2 / V4-Pro models require multi-node H100 / H200 / B200 deployments at full precision; quantized variants ship from the open-source community shortly after each release. On June 27, 2026 DeepSeek released DSpark — a speculative-decoding drafter that, per DeepSeek's own benchmarks, speeds up V4 per-user generation ~60–85% on V4-Flash and ~57–78% on V4-Pro with no change to the model — alongside DeepSpec, an MIT-licensed codebase for training and evaluating speculative-decoding draft models on open targets (Qwen3, Gemma). These are serving optimizations rather than new DeepSeek models, so neither is a table row; the numbers are DeepSeek-reported and not yet independently verified. As of the July 31, 2026 open-weights drop, DSpark is no longer a separate download for the current flagship — the official DeepSeek-V4-Flash-0731 checkpoint ships with the drafter built in, enabled in vLLM and SGLang with a single speculative-decoding flag and no separate draft-model path. The V4-Pro GA weights followed on August 13, 2026 under the same MIT terms, with a vLLM recipe linked from the model card. DeepSeek also open-sourced its own agent scaffold, DeepSeek Harness, the same day — the plugin-based framework its published agent benchmarks are run in. The multimodal V4-Flash-Vision-Exp weights landed on August 31, 2026, also MIT, with SGLang and vLLM recipes added to the card over the following day; that release broke the same-day pattern the two GA builds had set, trailing its API debut by ten days.

Hosted-inference providers. Together AI, Fireworks AI, OpenRouter, SiliconFlow, Groq, Perplexity's public-API tier. Most providers serve the post-MIT weights (R1 forward) and clearly label which version is hosted. This tier tracks DeepSeek closely: OpenRouter's catalog carries the entire V4 line — the April preview builds, the V4-Flash-0731 and V4-Pro-0813 GA builds, and V4-Flash-Vision-Exp — each listed within a day of its API release.

Hyperscalers. AWS Bedrock, Microsoft Azure AI Foundry, NVIDIA NIM, IBM watsonx, and Oracle OCI have all added DeepSeek SKUs across 2025–2026, and Google Cloud's Vertex AI Model Garden has tracked the line closely rather than lagging it — R1 and V3 in preview in February 2025, then V3-0324, V3.1, V3.1-Terminus, V3.2-Exp, DeepSeek-OCR, and V3.2 as each shipped, with V3.2 a fully-managed API since December 10, 2025. What no hyperscaler catalog carries yet is V4: checked on September 8, 2026, Bedrock's DeepSeek model cards top out at V3.2, Azure AI Foundry's featured-model list at V3-0324, and Vertex's release notes at V3.2 / V3.2-Speciale. On this family the hyperscalers run a release generation behind the hosted-inference providers; check the catalogs for the current state.

People who shaped DeepSeek

Liang Wenfeng — founder and CEO of DeepSeek, co-founder and CEO of High-Flyer. The 2021 GPU-stockpile decision, the July 2023 DeepSeek incorporation, the V2 MoE / MLA bet, and the 2025 MIT-licensing turn all trace through Liang's office. Profiled in Fortune; on Wikipedia at Liang Wenfeng.

High-Flyer (Hangzhou Huanfang Technology Co., Ltd.) — the parent quantitative hedge fund. Co-founded by Liang in February 2016; reported to be using AI exclusively for trading by 2021. The funder of the Fire-Flyer 2 cluster and DeepSeek's only investor until the first external round closed in mid-2026.

DeepSeek's research staff — the lab is known for an unusually flat structure, a young research team (many recent PhD graduates from Tsinghua, Peking University, and Zhejiang University), and a publication culture that ships technical reports alongside model releases. Named-author rosters appear on the V2 / V3 / R1 / V3.2 papers on arXiv. Several core researchers have reportedly been recruited away to ByteDance, Tencent, Xiaomi, and the autonomous-driving company Yuanrong Qihang during 2025; named-departure tracking is sparse compared to U.S. labs.

No publicly-named CTO, CEO-second, or board. Unlike OpenAI, Anthropic, xAI, and Google DeepMind, DeepSeek does not maintain a leadership page; corporate governance is held inside the High-Flyer / DeepSeek Liang-led structure. The 2026 first-external round was deliberately structured to preserve that concentration — investors hold limited-partnership interests with no voting rights, with only the state-backed National AI Industry Investment Fund taking a direct, voting stake — so it is unlikely to produce a conventional outside-investor board.

The competitive landscape

DeepSeek is, alongside Alibaba's Qwen line, one of the two dominant Chinese open-weights AI families through 2025–2026. The closest direct comparators on the open-weights axis are Alibaba's Qwen (also Apache-2.0-or-permissive across most releases, with strong HuggingFace-leaderboard presence — see Qwen Versions), Mistral (French; mixed Apache 2.0 / Mistral Research License / proprietary tiers across the line, with the December 2025 “Mistral 3” family relaunch re-committing the open releases to Apache 2.0 — see Mistral Versions), Meta's Llama (custom Llama Community License, see Llama Versions), and Moonshot AI's Kimi line. The closed-weights frontier competitors — ChatGPT, Claude, Gemini, Grok — are the practical benchmark for “is DeepSeek competitive at frontier scale,” which is the question the V3 / R1 / V3.2 / V4 release cycle has been answering in the affirmative since December 2024. This page does not attempt a benchmark roundup or a ranking.