DeepSeek

Updated June 2026

DeepSeek is the Hangzhou AI lab behind the DeepSeek app and the open-weight DeepSeek V4 and R-series models. Spun out of the quant hedge fund High-Flyer and run by founder Liang Wenfeng, it pairs frontier-adjacent quality with the lowest prices in the industry and an MIT licence, and in 2026 raised its first outside funding at a valuation above $50 billion.

Website →API →Pricing →

DeepSeek is the Chinese AI lab behind the DeepSeek app and the open-weight DeepSeek V4 and R-series models, and it is the company that proved frontier-adjacent AI could be built cheaply and given away. Founded in Hangzhou in July 2023 and spun out of the quantitative hedge fund High-Flyer, DeepSeek releases its models as open weights under the permissive MIT licence and undercuts every major lab on price — V4-Pro costs $0.87 per million output tokens, roughly 34 times cheaper than GPT-5.5. In 2026 it raised its first-ever outside funding, reportedly about $7.4 billion at a valuation above $50 billion, making it China’s most valuable AI start-up.

The strategy is the story. DeepSeek runs one of the smallest teams of any frontier-scale lab — on the order of a few hundred people — and competes not by matching the absolute frontier but by getting within three to six months of it at a fraction of the cost, then open-sourcing the result. That combination made its R1 model a global event in January 2025 and keeps DeepSeek central to the US-China AI rivalry, the open-weight movement, and a running set of disputes over data, distillation and privacy.

Quick facts

History and founding

DeepSeek was founded in July 2023 in Hangzhou by Liang Wenfeng, a quant investor who had built the hedge fund High-Flyer into one of China’s larger quantitative funds. High-Flyer had stockpiled Nvidia GPUs for its trading models, and DeepSeek began as a research bet that the same compute could be turned toward foundation models. For its first two years the lab took no outside capital at all — it was funded entirely off High-Flyer’s balance sheet, an unusual independence that let it publish openly and price aggressively without investor pressure.

DeepSeek became a global name in January 2025. Its DeepSeek-R1 reasoning model matched the quality of far more expensive Western systems, was released as open weights, and powered a free app that topped the US App Store. DeepSeek’s claim that R1’s lineage cost a fraction of a Western training run triggered a sharp sell-off in US technology stocks, with Nvidia alone losing close to $600 billion in market value in a single day. The “DeepSeek moment” reframed the AI race around efficiency rather than raw spend.

Funding and valuation

For most of its life DeepSeek was a rarity among frontier labs: profitable parent, no venture capital, no external valuation. That changed in 2026, when it closed its first outside funding round — reportedly about $7.4 billion at a valuation above $50 billion (Outlook Business). Estimates of the valuation vary across reports from roughly $45 billion to $59 billion; as a private Chinese company, DeepSeek publishes no audited financials, so these figures are press-sourced.

Two features make the round unusual. First, founder Liang Wenfeng committed the largest single slice personally — around 20 billion yuan (roughly $3 billion) — drawn from his High-Flyer wealth. Second, outside investors were reportedly required to route their money through a limited partnership managed by Liang rather than into the company directly, preserving founder control (HelloChinaTech). The biggest external backers were Tencent (around 10 billion yuan) and battery giant CATL (around 5 billion yuan), part of a wider Chinese push to fund a homegrown AI champion.

Even after the raise, the funding gap with Western rivals is stark. DeepSeek’s $50-billion-plus valuation is a fraction of Anthropic’s $965 billion or OpenAI’s figures — a reminder that DeepSeek competes on efficiency, not capital.

Models and the open-weight lineup

DeepSeek ships a compact lineup built around clear job-to-model matches, all under the MIT licence. The full set with live benchmarks renders below this page; the headline models are:

DeepSeek V4 is the current flagship, released as a preview on 24 April 2026 across web, app and API. It is a Mixture-of-Experts design in two sizes — V4-Pro at 1.6 trillion parameters (49 billion active) and V4-Flash at 284 billion (13 billion active) — both with a 1-million-token context window. V4 scores 80.6% on SWE-bench Verified (vendor-reported), the top open-weight coding score on the site’s best AI models board and within a fraction of a point of last year’s proprietary leaders.

On reasoning, the picture is more measured. DeepSeek R2 has not been released as of June 2026: CEO Liang Wenfeng is reportedly unsatisfied with its performance, and widely-circulated R2 specifications come from leaks, not the company. The DeepThink reasoning mode in the app still runs on the R-series. DeepSeek’s own V4 technical report concedes the model trails the absolute frontier, putting its reasoning and agentic ability roughly three to six months behind the current leaders — GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.8 (CFR).

The efficiency-and-price strategy

DeepSeek’s signature is not a training method but a business posture: frontier-adjacent quality, open weights, and the lowest prices in the industry. The clearest expression of that came in 2026, when DeepSeek made a 75% cut to V4-Pro API pricing permanent.

The discount began as a promotion. On 22 May 2026 DeepSeek announced it would not roll back — the discounted rate became the permanent list price. The company framed the cut not as a response to competition but as a pass-through of genuine efficiency: V4-Pro runs on a small fraction of the compute its predecessor needed, and DeepSeek chose to hand that saving to customers rather than bank it as margin (InfoWorld).

Cache hits are cheaper still, at roughly $0.0036 per million tokens. Combined with the MIT licence, the result is unusual: the same capability is available three ways — free in the app, near-free via the API, or self-hosted at no licence cost. For data-sensitive users, self-hosting is also the route around DeepSeek’s China data residency (see Controversies).

Products and ecosystem

• DeepSeek app — the free consumer chat across web, iOS and Android. At about 131.5 million monthly active users in January 2026 it is the second most-used AI assistant after ChatGPT.
• DeepSeek API — the developer platform and the company’s main revenue line, priced well below every major rival.
• Open weights — V4 and the R-series are published on Hugging Face under MIT, downloadable for self-hosting.
• Specialist models — Janus-Pro (multimodal) and DeepSeek-VL2 (vision and document understanding) extend the lineup beyond chat.

A defining 2026 development is hardware. R1 was trained on Nvidia GPUs; with V4, DeepSeek shifted its inference and serving onto domestic Chinese chips from Huawei (Ascend) and Cambricon, while V4’s pre-training still leaned on Nvidia H800s with Ascend used in parallel (CNN). It is a partial decoupling — a direct response to US export controls and a signal of where China’s AI stack is heading.

Business and financials

DeepSeek monetises through its low-cost API, not enterprise sales teams or consumer subscriptions — the app is free and has no paid tier. The model is volume at razor-thin prices, subsidised by the efficiency of its own architecture and, historically, by High-Flyer’s balance sheet. As a private company it discloses no audited revenue; the 2026 funding round is the first external mark on its value.

The headcount is the other half of the story. DeepSeek runs on a few hundred people at most — around 160 in 2025, with mid-2026 estimates up to roughly 300, still a fraction of OpenAI or Anthropic — with a young research-first culture and, by reports, little interest in scaling the organisation. It is one of the most capital- and labour-efficient operations in frontier AI, which is precisely what its pricing reflects.

Leadership

• Liang Wenfeng — founder and CEO; also founder of the quant fund High-Flyer, and the largest personal backer of DeepSeek’s 2026 round. He sets the lab’s research-led, open-weight direction and retains control through the round’s limited-partnership structure.

DeepSeek does not publicise a broad executive bench; it is unusually founder-centric, with research talent rather than named executives as its public face.

Competition and market position

DeepSeek’s position is distinct from the Western frontier labs. It does not try to hold the absolute top of the benchmark tables; it aims to be the best price-to-quality option and the leading open-weight lab, and it succeeds at both. On the site’s best AI models board, DeepSeek V4 leads the open-weight tier and sits within roughly 0.2 points of Gemini 3.1 Pro on SWE-bench Verified, at a fraction of the cost.

Its closest peers are the other Chinese open-weight labs — Alibaba’s Qwen, Moonshot’s Kimi and MiniMax — which occupy the same cluster. Against the frontier labs (OpenAI, Anthropic, Google), DeepSeek trades the last three-to-six months of capability for an order-of-magnitude price advantage and a licence that lets anyone run the model themselves. Its weakness is the mirror image: no absolute-frontier model, a reasoning successor (R2) still unreleased, and the trust and regulatory baggage that comes with being China-based.

Controversies

• Distillation allegations. In February 2026, OpenAI submitted a memo to the US Congress’s China Select Committee alleging DeepSeek distilled its models — training on ChatGPT outputs obtained through obfuscated third-party routers to mask their source. Days later, Anthropic accused DeepSeek, Moonshot and MiniMax of coordinated distillation against Claude. These are allegations; DeepSeek has not publicly conceded them, and distillation claims are difficult to prove.
• China data residency and government bans. The hosted app stores user data on servers in China, where the 2017 National Intelligence Law can compel cooperation with the state. Italy, Australia, South Korea, Taiwan, India, the Czech Republic and a number of US states and agencies have restricted or banned it, and Germany asked Apple and Google to delist it (Al Jazeera).
• The training-cost claim. DeepSeek’s widely-quoted figure that a flagship training run cost only about $5.6 million is contested: it covers a single final run and excludes prior research, failed runs and the cost of the GPU fleet, so it understates true development cost.
• Founder-control structure. The 2026 round’s requirement that investors route capital through a Liang-managed vehicle gives outside backers exposure without normal governance rights — efficient for the founder, unusual for investors of that size.

Recent developments (2026)

• DeepSeek V4 launched on 24 April 2026 (V4-Pro and V4-Flash, 1M context, MIT), with serving shifted onto domestic Huawei and Cambricon chips while pre-training still leaned on Nvidia.
• First outside funding (~$7.4B at $50B+, 2026) ended two years of pure self-funding and made DeepSeek China’s most valuable AI start-up.
• Permanent V4-Pro price cut (75%, made permanent 22 May 2026) escalated the global inference price war.
• Distillation disputes with OpenAI and Anthropic moved into the US policy arena.
• R2 remains unreleased, with the company reportedly holding it back rather than ship below its own bar.

Where DeepSeek excels

• Price-to-quality. The cheapest pricing of any major lab, with frontier-adjacent results — V4-Pro output is roughly 34x cheaper than GPT-5.5.
• Open weights. An MIT licence on a frontier-adjacent model makes self-hosting, fine-tuning and air-gapped deployment genuinely viable.
• Efficiency. A team in the low hundreds and a domestic-chip serving stack show a path to frontier-adjacent AI without frontier-scale capital.
• Reach. The free app is the world’s second most-used AI assistant, a vast distribution channel for the models.

Where DeepSeek falls short

• Trails the frontier. By its own admission, V4 sits three-to-six months behind the leading models, and the R2 reasoning successor is unreleased.
• Trust and compliance. China data residency and the distillation allegations make the hosted products a non-starter for many governments and enterprises.
• Thin governance. Founder-controlled structure and minimal corporate disclosure raise questions for risk-averse buyers.
• Capital gap. A $50-billion valuation is a fraction of its Western rivals’, constraining the compute it can throw at the next generation.

Developer resources

DeepSeek’s developer stack centres on the DeepSeek API (api-docs.deepseek.com), which is OpenAI-compatible and among the cheapest available, with prompt caching that further cuts cost. Current per-token rates live on the pricing page. For self-hosting, the open weights for V4 and the R-series are published under MIT on Hugging Face; for quality, serve in bf16 rather than fp8 quantisation. The models are also widely available through third-party inference providers and routers.

Frequently asked questions

What is DeepSeek?

DeepSeek is a Chinese AI lab, founded in Hangzhou in July 2023 and spun out of the quant hedge fund High-Flyer. It builds the open-weight DeepSeek V4 and R-series models and the free DeepSeek app, and is known for matching frontier-adjacent quality at the lowest prices in the industry.

Who owns DeepSeek?

DeepSeek is private and founder-controlled. CEO Liang Wenfeng, who also founded the hedge fund High-Flyer, is the largest backer and retains control even after the 2026 funding round, in which outside investors routed capital through a partnership he manages. Tencent and CATL are among the external backers.

How much is DeepSeek worth?

DeepSeek raised its first outside funding in 2026, reportedly about $7.4 billion at a valuation above $50 billion (estimates range from $45 billion to $59 billion), making it China’s most valuable AI start-up. As a private company it publishes no audited financials, so these are press-reported figures.

Is DeepSeek open source?

Yes. DeepSeek publishes its models, including DeepSeek V4 and the R-series, as open weights under the permissive MIT licence, so they can be downloaded, self-hosted, fine-tuned and used commercially. The app and API are separate paid-or-free services built on the same models.

Why is DeepSeek so cheap?

DeepSeek designs efficient Mixture-of-Experts models that run on a fraction of the compute of comparable systems, then passes the saving through to its API rather than banking it as margin. In May 2026 it made a 75% cut to V4-Pro pricing permanent. V4-Pro costs $0.87 per million output tokens, roughly 34 times cheaper than GPT-5.5.

Is DeepSeek safe to use?

The models themselves are standard open-weight models. The risk is the hosted app, which stores data on servers in China and has been restricted by several governments. For sensitive or regulated data, self-host the MIT-licensed weights or use a non-Chinese provider rather than the hosted app.

What is the latest DeepSeek model?

The current flagship is DeepSeek V4, released on 24 April 2026 in two sizes (V4-Pro and V4-Flash) with a 1-million-token context window. The R2 reasoning model has not been released as of June 2026.

Did DeepSeek copy OpenAI?

OpenAI and Anthropic have alleged that DeepSeek used distillation — training on their models’ outputs — to build its own. DeepSeek has not publicly conceded this, and such claims are hard to prove. It remains a contested allegation rather than an established fact.

Models

Apps

• DeepSeek