DeepSeek Company Profile

The CODEW · Company Profile · AI Intelligence

Last updated: October 20, 2026 | Latest model release: DeepSeek-V4.1-Flash-1017 (October 17, 2026) | Ownership: Private; founder-controlled

Company overview

DeepSeek is a Hangzhou-based artificial-intelligence laboratory that develops frontier large language models and releases many of them as open-weight models under permissive licenses. It was founded in July 2023 by Liang Wenfeng as a spin-off from High-Flyer, the Chinese quantitative hedge fund he co-founded, and it became globally prominent in early 2025 after demonstrating that highly capable reasoning models could be trained and served at dramatically lower cost than many Western competitors.

The company’s commercial model combines a low-priced API, a consumer chat app, and freely downloadable model weights. That strategy has produced rapid adoption: DeepSeek’s annualized revenue run rate reportedly reached $1 billion in September 2026, more than double the level reported a few months earlier, while its latest V4.1-Flash-1017 release continues to compete on price, speed, coding, and agentic performance.

Key facts

Founded	July 2023
Headquarters	Hangzhou, China
Founder and CEO	Liang Wenfeng, co-founder of High-Flyer
Ownership	Private; Liang Wenfeng retains an estimated stake of roughly 78% after the 2026 funding round
Funding	First major external round closed in June 2026: about $7.4 billion at a post-money valuation near $52 billion; a further round of about $7.5 billion was reportedly being finalized at a valuation near $74 billion
Revenue	About $1 billion annualized run rate as of September 2026; roughly $70.7 million of recognized revenue in January–July 2026
Core business	Frontier AI models, developer API, consumer chat assistant, and open-weight model distribution
Licensing	Many flagship models released with open weights under the MIT License

Business model

DeepSeek monetizes its research primarily through pay-per-token API access, while using open-weight releases to broaden distribution, developer trust, and ecosystem adoption. Developers can either call DeepSeek’s hosted API or download eligible model weights and run them on their own infrastructure, a flexibility that differentiates it from closed-model rivals such as OpenAI and Anthropic.

The company’s pricing is deliberately aggressive. For DeepSeek-V4.1-Flash-1017, published API rates are $0.30 per million uncached input tokens and $1.20 per million output tokens during peak hours, with both rates halved during off-peak periods; cached input is priced far lower. DeepSeek has also raised prices on some models while retaining demand, which helped lift its annualized revenue pace toward $1 billion.

Products and technology

Area	Status
DeepSeek-V4 family	DeepSeek-V4 was released in April 2026, followed by V4 Preview with a 1 million-token context window in June 2026. The family targets reasoning, coding, multimodal tasks, and long-context applications.
V4.1-Flash	A smaller, faster model line launched in September 2026, designed for lower-cost inference and quicker task completion.
V4.1-Flash-1017	Released October 17, 2026. DeepSeek reported a Humanity’s Last Exam score of 42.6 without tools, up from 36.8, and a Terminal-Bench 4.0 coding-agent score of 42.3, up from 31.2.
Open weights	Flagship models are commonly released under the MIT License, allowing commercial use, self-hosting, fine-tuning, and redistribution subject to the license terms.
Developer platform	Hosted API with peak and off-peak pricing, caching, and model access through DeepSeek’s website, app, and developer platform.

Financial information

Metric	Latest reported / estimated
Annualized revenue run rate	About $1 billion as of September 2026, according to reporting citing people with direct knowledge
Recognized revenue	About 475 million yuan, or $70.7 million, in January–July 2026
Funding	About $7.4 billion raised in June 2026; Liang Wenfeng personally committed roughly $3 billion
Valuation	About $52 billion post-money after the June round; later reports indicated a proposed follow-on round near $74 billion
API pricing	V4.1-Flash-1017: $0.30 per million uncached input tokens and $1.20 per million output tokens at peak; $0.15 and $0.60 off-peak
Potential listing	Reports indicate preparation for a potential Shanghai listing, though timing and valuation remain undecided

DeepSeek does not publish audited quarterly financial statements comparable with a listed company, so its revenue and valuation figures should be treated as press and investor-report estimates rather than confirmed public-company disclosures. The sharp revenue increase followed substantial API price increases—reported at roughly 2.3 to 4.5 times prior levels—without an equivalent collapse in demand.

Leadership

Liang Wenfeng

Founder and CEO; co-founder of High-Flyer. He remains DeepSeek’s dominant shareholder and has personally invested heavily in the company’s recent funding rounds.

DeepSeek discloses less executive detail than U.S.-listed AI companies. Confirm current research and product leadership through its official channels before publication.

Competitors

Competitor	Position versus DeepSeek
OpenAI	Leads in brand, enterprise distribution, and closed-model products; DeepSeek competes mainly on price, open weights, and cost-efficient reasoning.
Anthropic	Strong in enterprise coding and safety-oriented deployments; DeepSeek’s latest releases target coding and agentic workloads directly.
Google DeepMind	Competes through Gemini models, cloud distribution, and multimodal research.
Alibaba / Qwen	Major Chinese open-weight rival with broad model family and cloud integration; DeepSeek’s releases have pressured Chinese AI stocks, including Alibaba.
Moonshot AI, Z.ai, MiniMax	Domestic Chinese competitors in frontier and low-cost models; DeepSeek’s September V4.1-Flash launch reportedly pressured shares of MiniMax, Z.ai, and Alibaba.
Meta	Competes in the open-weight ecosystem through Llama, though with a different commercial strategy centered on platform and advertising.

Strategy

1. Efficiency as a product advantage

DeepSeek’s central thesis is that frontier capability does not require the highest possible infrastructure cost. Its models emphasize efficient training, compact deployment, fast inference, and low token pricing, making them attractive to developers and cost-sensitive enterprises.

2. Open weights with hosted monetization

By releasing weights under MIT-style terms, DeepSeek creates a large developer funnel while retaining a paid hosted API for users who do not want to operate models themselves. This mirrors an infrastructure-software playbook: free adoption at the edge, monetization through managed services.

3. Rapid release cadence

The company has moved from V4 in April to V4 Preview in June, V4.1-Flash in September, and V4.1-Flash-1017 in October. The cadence is designed to maintain benchmark relevance and respond quickly to competition from both Chinese and U.S. labs.

4. Capitalize without surrendering control

DeepSeek’s 2026 funding structure reportedly preserved founder control through vehicles and non-voting arrangements, while bringing in investors including Tencent, CATL, JD.com, NetEase, and a state-backed AI fund. A potential Shanghai IPO could provide a future liquidity path.

The CODEW analysis

Independent analysis, separate from the factual record above. Draft for editorial review.

DeepSeek has turned cost into a strategic weapon. Its pricing is low enough to make experimentation inexpensive, yet the company has shown that it can raise prices substantially and still grow revenue. That suggests real demand elasticity is more favorable than a simple “cheap Chinese model” narrative implies.

Open weights are both its moat and its constraint. MIT-licensed weights make DeepSeek easy to adopt, but they also allow rivals, cloud providers, and enterprises to self-host the technology without paying DeepSeek. The durable monetization question is whether its hosted API, latency, tooling, and model updates can keep enough usage on DeepSeek’s own platform.

Compute availability is a visible operational risk. Service slowdowns and an outage around the V4.1-Flash-1017 launch were attributed by some observers to demand outpacing computing resources. For an AI company whose advantage is low-cost, high-volume inference, reliability is now as important as benchmark scores.

Geopolitics limits the addressable market. DeepSeek can be downloaded and self-hosted, but use of its China-hosted services can raise export-control and data-governance issues for some U.S. organizations. This does not eliminate Western adoption, but it pushes sensitive enterprise workloads toward local deployment or non-Chinese providers.

What to watch: whether the reported $1 billion run rate converts into sustained recognized revenue; pricing changes and gross-margin disclosure; compute capacity and service reliability; V4-family enterprise adoption; progress toward a Shanghai listing; and whether open-weight rivals erode DeepSeek’s benchmark lead.

Related coverage

AI Intelligence
Startup Funding Watch: DeepSeek’s $12B Round, Lambda’s $4B Pre-IPO and the AI Capital Chain
Open Source Watch: OpenTelemetry's CNCF Milestone, DeepSeek V4 Flash & OpenTofu Governance
IPO Watch: DeepSeek Seeks Shanghai IPO, Moonshot AI Confidentially Files for Hong Kong Listing, AI IPO Wave Faces Public-Market Test
AI Watch: Anthropic Fund, Microsoft's Nscale Exit, Copilot Billing, DeepSeek's Mega Round and Nvidia's China Share

Sources

This profile is a durable reference page and is not investment advice. Factual sections are sourced above; the CODEW analysis section reflects The CODEW's independent view. Last updated October 20, 2026.

The CODEW