DeepSeek
DeepSeek is a Hangzhou, China-based AI company that builds open-source large language models (V4-Pro and V4-Flash MoE, MIT-licensed), serving developers and enterprises globally through an OpenAI/Anthropic-compatible API, free consumer web/app, and embedded integrations with Tencent, Huawei, and Microsoft.
- Company typePrivate
- Founded2023
- HeadquartersHangzhou, China
- Headcount101–250
- GTM typeB2B
- OfferingSoftware
What DeepSeek does
DeepSeek (杭州深度求索人工智能基础技术研究有限公司 / Hangzhou DeepSeek Artificial Intelligence Co., Ltd.) is a privately held Chinese AI research and product company founded in 2023 in Hangzhou by Liang Wenfeng, also the founder of the High-Flyer quantitative hedge fund. The company builds open-source large language models (LLMs) based on deep neural networks trained via large-scale self-supervised pre-training and targeted optimization (SFT and reinforcement learning), with all model weights, parameters, and inference tool code released under the permissive MIT License. Its current flagship model family is DeepSeek-V4 (Preview released April 2026), comprising DeepSeek-V4-Pro (1.6T total / 49B active parameters MoE) and DeepSeek-V4-Flash (284B total / 13B active), both supporting a 1M-token context window, hybrid thinking/non-thinking modes, and novel architectural innovations including DeepSeek Sparse Attention (DSA) and Manifold-Constrained Hyper-Connections (mHC). Previous-generation DeepSeek-V3.2 and DeepSeek-R1 models remain accessible via API aliases.
The company serves developers and enterprises globally through a product stack that includes a free consumer web chat (chat.deepseek.com) and mobile app, a self-serve developer API platform (platform.deepseek.com) offering OpenAI ChatCompletions- and Anthropic-compatible endpoints, open-source distribution on GitHub and HuggingFace, and embedded integrations inside major platforms including Tencent WeChat (1.4B users) and WeCom, Huawei Cloud, Zoho SalesIQ, and 20+ Chinese automakers. Revenue is generated primarily through pay-as-you-go API token consumption (per million input/output tokens, with separate cache-hit and cache-miss pricing in CNY or USD), supplemented by indirect enterprise revenue via hosting partners such as SiliconFlow and SAIHEAT.
In June 2026, DeepSeek closed its first external funding round of approximately $7.4 billion at a valuation exceeding $50 billion — China's largest AI financing — from a syndicate including Tencent, CATL, JD.com, NetEase, IDG Capital, and China's National AI Industry Investment Fund, with founder Liang Wenfeng personally contributing ~$3B. The deal was structured through a founder-controlled limited partnership with five-year lock-ups and no voting rights for commercial investors, preserving absolute strategic control. The company has ~150 employees (protected by a no-poaching clause) and announced plans to double all technical and engineering departments, while operating under PRC generative AI service and CAC algorithm registrations that align it with Beijing's sovereign AI priorities.
DeepSeek firmographics
Firmographics- Name
- DeepSeek
- Legal name
- Hangzhou DeepSeek Artificial Intelligence Co., Ltd. (杭州深度求索人工智能基础技术研究有限公司)
- Website
- https://deepseek.com
- Company type
- Private
- Founded year
- 2023
- Operating status
- Operating
- Headcount range
- 101–250 employees
- Short description
- DeepSeek is a Hangzhou, China-based AI company that builds open-source large language models (V4-Pro and V4-Flash MoE, MIT-licensed), serving developers and enterprises globally through an OpenAI/Anthropic-compatible API, free consumer web/app, and embedded integrations with Tencent, Huawei, and Microsoft.
- Ownership category
- akta.pro rank
DeepSeek industry classification
Industry- Product category
- Foundation Model API
- NAICS
- Computer Systems Design Services (541512)
- akta.pro primary industry
- Foundation Model Developers (LLM/Multimodal Model Labs) (HDAAACAA)
- akta.pro secondary industries
- Fine-Tuning, Adaptation & Custom Model Training (PEFT/LoRA/RLHF) (HDAAACAC), NLP Developer Tools & Evaluation (prompting, testing, benchmarks) (HDAAADAK)
Keywords
Where DeepSeek is headquartered
LocationHeadquarters
- HQ city
- Hangzhou
- HQ country
- China
- HQ region
- Asia
Offices1 record
Markets served
DeepSeek business model
Business model- GTM type
- B2B
- Offering type
- Software
- Cost components
- Technology or R&D, Personnel, Infrastructure, Operations, Marketing or Sales
Revenue model
- API Token Usage (Pay-as-you-go): Primary revenue stream: developers and enterprises pay per million input/output tokens consumed via the OpenAI- and Anthropic-compatible API. Two SKUs — deepseek-v4-flash (cost-efficient) and deepseek-v4-pro (flagship) — billed in CNY or USD with separate cache-hit (discounted) and cache-miss pricing. Users top up via PayPal, bank card, Alipay, or WeChat Pay on the self-serve platform.
- Free Consumer Chat (Direct-to-Consumer): Free access to DeepSeek models via chat.deepseek.com and DeepSeek App, serving as a top-of-funnel acquisition channel and brand-building tool rather than a direct revenue stream.
- Enterprise/B2B Platform Partnerships: Indirect enterprise revenue via hosting partners (SiliconFlow, Huawei Cloud, SAIHEAT, Zoho) and embedded integrations (Tencent WeCom Dayuan, WeChat Xiaowei) that consume DeepSeek models at scale and pay for API access or licensing.
Pricing tiers
| Model | Billing | Price |
|---|---|---|
| Unit Pricing | Pay-as-you-go | DeepSeek-V4-Flash: low-cost MoE with 1M context, 13B active params |
| Unit Pricing | Pay-as-you-go | DeepSeek-V4-Pro: flagship 1.6T MoE with 49B active params |
| Freemium | Pay-as-you-go | Free consumer tier via web chat and App |
Go-to-market motion1 record
Distribution channels7 records
Marketing channels10 records
DeepSeek product offering
Product offeringCore offering
DeepSeek develops and open-sources large language models built on deep neural networks via large-scale self-supervised pre-training and targeted optimization (SFT and RL). Its flagship DeepSeek-V4 family — V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B total / 13B active parameters) — uses a Mixture-of-Experts architecture with DeepSeek Sparse Attention (DSA) for 1M-token context efficiency. Models are monetized primarily through a pay-as-you-go OpenAI- and Anthropic-compatible API at platform.deepseek.com, with a free consumer web (chat.deepseek.com) and mobile app tier, and all model weights released under the MIT License.
Product overview
DeepSeek offers a single unified platform of generative-AI products, built around three consumer-facing entry points—DeepSeek App (mobile), DeepSeek Chat (web), and the DeepSeek Platform / Open Platform (developer API)—all powered by a shared family of self-developed, open-source large language models. The current flagship model family is DeepSeek-V4 (released April 24, 2026 in Preview), comprising DeepSeek-V4-Pro (1.6T/49B-active) for top-tier reasoning and agentic coding, and DeepSeek-V4-Flash (284B/13B-active) for fast, economical inference; both expose Thinking and Non-Thinking modes over a 1M-token context and integrate with Claude Code, GitHub Copilot, OpenCode, and LangChain. Previous-generation models DeepSeek-V3.2 and DeepSeek-R1 (R1-0528) remain accessible via the API as deepseek-chat/deepseek-reasoner aliases, all open-sourced under the permissive MIT License. The DeepSeek-V4 family introduces DeepSeek Sparse Attention (DSA) for ultra-low-cost 1M-context inference, supports Function Calling, JSON Output, FIM completion, Chat Prefix Completion, and Context Caching on Disk, and is positioned as the engine behind DeepSeek's in-house Harness team and the forthcoming "DeepSeek Code" agentic product.
Differentiator
Problem solved
Functional benefit
Brands
- DeepSeek App: DeepSeek's mobile AI assistant application available on iOS and Android platforms.
- DeepSeek Chat
- DeepSeek Platform (Open Platform)
Products and services
- DeepSeek App Official mobile application for DeepSeek, available on iOS and Android, providing free chat-based access to the latest DeepSeek models (V4-Pro, V4-Flash, V3.2, R1). Supports voice input, camera/image upload, multi-turn conversations, image recognition, search-augmented generation, and account registration/login via phone number, one-click carrier login, or third-party login (WeChat, Apple, Google).
- DeepSeek Chat (Web) Browser-based chat interface at chat.deepseek.com, providing free conversational access to DeepSeek's flagship models (DeepSeek-V4 via Expert Mode / Instant Mode, plus DeepSeek-V3.2 and DeepSeek-R1). Supports Expert Mode (thinking/reasoning), Instant Mode, conversation sharing, history management, image recognition, and search-augmented generation.
- DeepSeek Platform (Open Platform) Commercial API/developer platform at platform.deepseek.com providing programmatic access to DeepSeek's models via OpenAI ChatCompletions and Anthropic API-compatible endpoints. Includes API key management, top-up/billing in CNY or USD via PayPal, bank card, Alipay, WeChat Pay, usage analytics by API key, account management, refund processing, and identity verification (individual and enterprise).
- DeepSeek-V4-Pro Flagship large language model with 1.6T total / 49B active parameters; performance rivals top closed-source models. Supports Thinking and Non-Thinking modes, 1M token context length, up to 384K output, JSON Output, Function Calling, Chat Prefix Completion (Beta), and FIM (Beta, non-thinking). Open-source under MIT License. Pricing: 0.025 RMB/M tokens input (cache hit), 3 RMB/M input (cache miss), 6 RMB/M output. Concurrency limit 500.
- DeepSeek-V4-Flash Efficient, cost-effective DeepSeek-V4 variant with 284B total / 13B active parameters; reasoning capability approaches V4-Pro while delivering smaller-parameter, faster responses. Supports 1M context, up to 384K output, all of V4-Pro's API features. Open-source under MIT License. Pricing: 0.02 RMB/M input (cache hit), 1 RMB/M input (cache miss), 2 RMB/M output. Concurrency limit 2500. As of June 2026, it was the #1 most-used model on OpenRouter globally with 4.63 trillion tokens consumed.
- DeepSeek-V3.2 DeepSeek's previous-generation hybrid reasoning model released December 1, 2025, exposed via deepseek-chat (non-thinking) and deepseek-reasoner (thinking). Released to Hugging Face with model card and technical report; succeeded in product line by DeepSeek-V4 in April 2026.
- DeepSeek-R1 (R1-0528) DeepSeek's flagship reasoning model released January 20, 2025 (upgraded to R1-0528 on May 28, 2025), exposed via deepseek-reasoner. Achieves OpenAI o1-comparable performance on math, code, and natural-language reasoning via large-scale reinforcement learning post-training. Open-sourced under MIT License; six distilled smaller models (1.5B–70B) released alongside; outputs both reasoning_content and final answer.
Quantifiable outcome
- 10x–50x lower cost per token than U.S. frontier models (by April 2026)
- +7 more outcomes
Companies that use DeepSeek
Customer profileNamed customers9 records
Ideal customer profiles3 records
DeepSeek technology and API
TechnologyTechnology focussed Yes
API detail
- Has API
- Yes
- API docs
- API detail
Core technology
AI maturity
App detail
Integration8 records
AI capability12 records
Feature9 records
DeepSeek partnerships and signals
Strategic signalPartnerships
Nine partnerships are on record, tiered core, flagship and minor.
- SiliconFlowcoreAI infrastructure startup hosting and optimizing DeepSeek models alongside Qwen, Zhipu and Moonshot on a multi-model, multi-chip platform serving over 10 million users and 10,000 enterprise customers; raised ~$300M in fifth funding round with Alibaba Cloud and Zhipu as earlier backers.
- MicrosoftflagshipMicrosoft is evaluating a fine-tuned version of DeepSeek V4 as a lower-cost, self-hosted open-source engine for its Copilot Cowork enterprise AI product. Microsoft Research found agentic coding tasks consume ~1,000x more tokens than standard queries, making DeepSeek's cost economics highly attractive. Adoption faces political headwinds from U.S. export controls and proposed No DeepSeek on Government Devices Act.
- Zoho (SalesIQ)coreZoho SalesIQ's Summer '26 update integrated DeepSeek alongside Anthropic and Google AI in its Agentic Intelligence platform, enabling enterprises to swap or combine models for compliance and flexibility.
- SAIHEAT (NASDAQ: SAIH)minorSAIHEAT expanded into AI inference services providing enterprise-level authorized token access to DeepSeek (alongside Kimi, GLM, MiniMax, MiMo) on dedicated high-performance AI infrastructure.
- Huawei TechnologiesflagshipHuawei-led research group (with Shenzhen Loop Area Institute, Harbin Institute of Technology Shenzhen, and Shenzhen Research Institute of Big Data) successfully completed full-parameter post-training of DeepSeek-V4-Pro (1.6T params) on a cluster of 1,000+ Huawei Ascend 910C chips in Shenzhen, demonstrating Chinese AI accelerator capability. DeepSeek also joined Huawei Cloud's AI Model Partner Program at INSPIRE 2026, alongside 20+ model providers.
- Cui Tianyi (formerly Jane Street)coreFormer Jane Street quantitative trading expert joined DeepSeek in March 2026 to lead the newly formed Harness team, focused on code agent products (potentially "DeepSeek Code") benchmarking against Anthropic's Claude Code.
- Geely, Chery, Dongfeng and 20+ Chinese automakerscoreDeepSeek models integrated by over 20 Chinese automakers (including Geely, Chery, Dongfeng) into in-vehicle AI assistants in early 2025, marking a fundamental shift from buyer-seller relationships to ecosystem-level AI/automaker partnerships.
- Anthropic (Claude Code)coreDeepSeek V4-Pro and V4-Flash are seamlessly integrated with Anthropic's Claude Code AI agent via Anthropic-compatible API base_url (api.deepseek.com/anthropic), enabling drop-in substitution without code changes.
- OpenAIcoreDeepSeek API uses an OpenAI ChatCompletions-compatible format at api.deepseek.com, enabling drop-in use with the OpenAI SDK and any OpenAI-compatible agent/coding tools.
Scale indicators12 records
Recent moves7 records
Expansion highlights7 records
DeepSeek competitors and assessment
Company assessmentDirect peers
- OpenAI: Frontier closed-source LLM provider (GPT-5/o-series) and DeepSeek's primary competitor in global enterprise and developer mindshare. DeepSeek's V4-Pro explicitly benchmarks against OpenAI frontier models, and DeepSeek-R1 was positioned as 'on par with OpenAI o1' at launch.
- Anthropic: Frontier LLM provider (Claude 4.5 series) and direct competitor in enterprise agentic coding. DeepSeek's API is Anthropic-compatible specifically to enable drop-in Claude Code substitution, and DeepSeek is currently building a 'DeepSeek Code' product benchmarked against Claude Code.
- Alibaba Qwen: Closest Chinese open-weight LLM competitor; Qwen3 series competes head-to-head with DeepSeek-V3/V4 on HuggingFace, OpenRouter, and global developer share. Alibaba Cloud is also a major Chinese cloud provider competing with DeepSeek's hosting partners.
- Mistral AI: European open-weight LLM provider (Mistral Large, Mixtral) with similar MIT-style licensing and enterprise API offering. Competes with DeepSeek for European enterprise customers seeking open-weight alternatives to U.S. closed models.
- Moonshot AI (Kimi): Chinese frontier LLM lab (Kimi K2) competing in the same Chinese open-weight LLM tier; named alongside DeepSeek in Anthropic's distillation accusations. Co-occurs with DeepSeek on OpenRouter global rankings and in SAIHEAT's inference offerings.
- Zhipu AI (Z.ai): Chinese frontier LLM lab (GLM-4.5/4.6) offering open-weight models that compete directly with DeepSeek on benchmarks, OpenRouter share, and via co-hosting on SiliconFlow. Also named alongside DeepSeek in Anthropic's distillation accusations.
Broad incumbents
- Meta (Llama): Open-weight LLM incumbent (Llama 4 series) with vastly larger distribution and brand. While not a direct frontier-reasoning competitor to V4-Pro, Llama defines the open-source LLM category and competes for the same self-hosted enterprise deployments.
- Google DeepMind: Frontier closed-source LLM provider (Gemini 2.5/3) with full-stack infrastructure (TPU, Vertex AI). Competes with DeepSeek for global enterprise spend, and Microsoft is reportedly considering DeepSeek V4 partly as a hedge against rising Gemini and OpenAI costs.
Others
- SiliconFlow: AI inference platform that hosts DeepSeek alongside Qwen, Zhipu, and Moonshot, serving 10M+ users and 10,000+ enterprise customers. Functions as a channel/distribution partner for DeepSeek rather than a direct competitor, but also competes indirectly as a multi-model platform.
Emerging players
- xAI: Frontier LLM lab (Grok 4) with strong X/Twitter distribution and rapidly scaling compute. Competes with DeepSeek in the open-weight-leaning frontier tier and for developer mindshare on cost-per-token benchmarks.
Market position
Strengths4 records
Weaknesses5 records
Competitive moat7 records
Key risks6 records
Key highlights7 records
Customer concentration
DeepSeek social profiles
Digital presenceDeepSeek compliance and trust
Trust signalCompliance5 records
DeepSeek financial estimates
Financial estimateRevenue estimate
Valuation estimate
DeepSeek leadership team
Management profileNumber of profiles
Profiles5 records
DeepSeek funding detail
Funding detailFunding overview
Funding rounds1 record
Investors8 records
Funding detail is available on the Subscription and Enterprise plan.Contact sales →
DeepSeek M&A and investment
M&A and investmentM&A
Investments
M&A and investment is available on the Subscription and Enterprise plan.Contact sales →
Frequently asked questions about DeepSeek
What does DeepSeek do?
DeepSeek develops and open-sources large language models built on deep neural networks via large-scale self-supervised pre-training and targeted optimization (SFT and RL). Its flagship DeepSeek-V4 family — V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B total / 13B active parameters) — uses a Mixture-of-Experts architecture with DeepSeek Sparse Attention (DSA) for 1M-token context efficiency. Models are monetized primarily through a pay-as-you-go OpenAI- and Anthropic-compatible API at platform.deepseek.com, with a free consumer web (chat.deepseek.com) and mobile app tier, and all model weights released under the MIT License.
Is DeepSeek a public or private company?
DeepSeek is a private company. It is classified as venture growth investor backed and is currently operating.
When was DeepSeek founded?
DeepSeek was founded in 2023. It employs 101 to 250 people.
Where is DeepSeek based?
DeepSeek is headquartered in Hangzhou, China, in the Asia region.
How does DeepSeek make money?
Three revenue lines are on record. API Token Usage (Pay-as-you-go) is the primary driver. The others are free Consumer Chat (Direct-to-Consumer) and enterprise/B2B Platform Partnerships.
Who are DeepSeek's main competitors?
Direct peers on record are OpenAI, Anthropic, Alibaba Qwen, Mistral AI, Moonshot AI (Kimi) and Zhipu AI (Z.ai). Broad incumbents are Meta (Llama) and Google DeepMind. SiliconFlow is listed as an others. xAI is listed as an emerging player.
Does DeepSeek have an API?
Yes. DeepSeek provides a public REST API compatible with the OpenAI ChatCompletions format (base_url: https://api.deepseek.com) and the Anthropic API format (base_url: https://api.deepseek.com/anthropic). The API supports chat completions, multi-round conversations, thinking/non-thinking mode switching, JSON Output, Function Calling / Tool Calls (up to 128 functions, with parallel calls), Chat Prefix Completion (Beta), FIM Completion (Beta), streaming (SSE) and non-streaming responses, context caching on disk (cache hit at $0.014/M tokens), user_id based rate limiting and isolation, and reasoning_effort control (high/max). Authentication is via Bearer API key issued at platform.deepseek.com. The DeepSeek API is explicitly supported as a backend by Claude Code, GitHub Copilot, OpenCode, and LangChain. SDK clients include the official OpenAI Python and Node.js clients; a sample deepseek_langchain.py is provided. Developer documentation is at api-docs.deepseek.com.
What industry is DeepSeek in?
DeepSeek's product category is Foundation Model API. Its primary akta.pro industry code is HDAAACAA, Foundation Model Developers (LLM/Multimodal Model Labs), with a secondary code of HDAAACAC, Fine-Tuning, Adaptation & Custom Model Training (PEFT/LoRA/RLHF). Its NAICS code is 541512.