# DeepSeek

> DeepSeek is a Hangzhou AI research company that builds large language and reasoning models, distributes downloadable weights under permissive licenses, and sells metered access through its API. Its breakthrough was as much economic as technical: a sparse model architecture, low-level systems engineering and reinforcement learning produced competitive results while using less compute per task, forcing the rest of the market to reconsider what frontier AI should cost.

- **Founded:** 2023
- **Headquarters:** Hangzhou, China
- **Founders:** Liang Wenfeng (Founder and CEO)
- **Products:** DeepSeek LLM and Coder, DeepSeek-V2, DeepSeek-V3, DeepSeek-R1, DeepSeek Chat
- **Notable:** Published DeepSeek-V3 with 671 billion total parameters but 37 billion activated per token, combining mixture-of-experts routing, Multi-head Latent Attention and FP8 training., Reported 2.788 million H800 GPU-hours for V3's full training run, estimated at $5.576 million at $2 per GPU-hour; the figure excludes prior research, experiments, data and infrastructure., Released R1 and six distilled variants under permissive licenses, enabling commercial use, modification and derivative work subject to base-model terms.

## Products & services

- **DeepSeek LLM and Coder** — The company's first general-language and code-focused open model families.
- **DeepSeek-V2** — A mixture-of-experts model that introduced DeepSeek's economical inference positioning and helped trigger a model-price reset in China.
- **DeepSeek-V3** — A 671-billion-parameter mixture-of-experts model with 37 billion parameters activated per token, trained using FP8 and hardware-aware systems optimizations.
- **DeepSeek-R1** — A reasoning model trained with reinforcement learning and multi-stage post-training, released with downloadable weights and smaller distilled variants.
- **DeepSeek Chat** — A free web and mobile assistant for writing, search, coding and multi-step problem solving.
- **DeepSeek API** — OpenAI-compatible, usage-priced hosted access for developers building applications on DeepSeek models.
- **DeepSeek-V3.2** — An open model update focused on agent capabilities, tool use and integrated reasoning.
- **DeepSeek-V4 family** — A million-token-context model family with Pro and Flash variants, followed by V4.1-Flash with native visual understanding.

## Achievements

- Published DeepSeek-V3 with 671 billion total parameters but 37 billion activated per token, combining mixture-of-experts routing, Multi-head Latent Attention and FP8 training.
- Reported 2.788 million H800 GPU-hours for V3's full training run, estimated at $5.576 million at $2 per GPU-hour; the figure excludes prior research, experiments, data and infrastructure.
- Released R1 and six distilled variants under permissive licenses, enabling commercial use, modification and derivative work subject to base-model terms.
- Reached No. 1 on the US Apple App Store's free-app chart in January 2025.
- Prompted a broad reassessment of AI infrastructure economics; Nvidia fell 16.9% on January 27, 2025, losing about $593 billion in market value in one day.
- Turned R1-Zero's readability and language-mixing failure into a stronger multi-stage R1 training recipe using cold-start data and a language-consistency reward.
- Established global distribution through open repositories and managed offerings from AWS and Microsoft.

## Latest updates

- **2025-11** — DeepSeek released V3.2, adding stronger agent and tool-use behavior while preserving an open-model distribution path.
- **2026-01** — The company published Engram and DeepSeek-OCR 2 research, extending its work into scalable memory and visual causal flow.
- **2026-06** — DeepSeek introduced V4, centered on efficient million-token context intelligence.
- **2026-08** — DeepSeek made V4-Pro generally available through its API.
- **2026-09** — DeepSeek announced V4.1-Flash, a smaller model in the V4 architecture family with native visual understanding.

## Links

- Website: https://www.deepseek.com/
- LinkedIn: https://www.linkedin.com/company/deepseek-ai
- Twitter/X: https://x.com/deepseek_ai
- GitHub: https://github.com/deepseek-ai

---

Profile page: https://yespress.io/deepseek
Published by YesPress — https://yespress.io
Last updated: 2026-09-26