OpenModelsHQ

DeepSeek

Hangzhou, China · High-Flyer · China Desk

DeepSeek is based in Hangzhou, China and is part of High-Flyer. We track 20 open-weight models from it, and the newest is DeepSeek-V4.1-Flash-1017, released 17 Oct 2026.

Models tracked
20
LLM 15 · Vision 5
Newest release
DeepSeek-V4.1-Flash-1017
17 Oct 2026
Country
China
Releases timeline
20 DeepSeek releases from 25 Jan 2024 to 17 Oct 2026. The largest is DeepSeek-V4-Pro-0813 at 1.65T parameters. Dot area follows total parameters; a ring means the size is not published.
Releases timeline: data
ModelCategoryReleasedParameters
DeepSeek-V4.1-Flash-1017Vision17 Oct 2026763B
DeepSeek-V4.1-FlashVision10 Sep 2026763B
DeepSeek-V4-Flash-Vision-ExpVision31 Aug 2026305B
DeepSeek-V4-Pro-0813LLM13 Aug 20261.65T
DeepSeek-V4-Flash-0731LLM31 Jul 2026304B
DeepSeek-V4-Flash-DSparkLLM27 Jun 2026304B
DeepSeek-V4-ProLLM22 Apr 20261.6T
DeepSeek-V4-FlashLLM22 Apr 2026291B
DeepSeek-OCR-2Vision27 Jan 20263.4B
DeepSeek-V3.2LLM1 Dec 2025685B
DeepSeek-OCRVision17 Oct 20253.3B
DeepSeek-R1-0528-Qwen3-8BLLM29 May 20258.2B
DeepSeek-V3-0324LLM24 Mar 2025685B
DeepSeek-R1-Distill-Qwen-32BLLM20 Jan 202532.8B
DeepSeek-R1-Distill-Qwen-7BLLM20 Jan 20257.6B
DeepSeek-R1-Distill-Qwen-1.5BLLM20 Jan 20251.8B
DeepSeek-R1LLM20 Jan 2025684B
DeepSeek-V3LLM25 Dec 2024684B
DeepSeek-Coder-V2-Lite-InstructLLM14 Jun 202415.7B
deepseek-coder-7b-instruct-v1.5LLM25 Jan 20246.9B

Origin

Liang Wenfeng founded DeepSeek on 17 July 2023 as a spin-off from High-Flyer, the hedge fund he co-founded in June 2015. High-Flyer owns and funds it. The company is headquartered in Hangzhou, Zhejiang.

How it got here

DeepSeek-Coder came first, on 2 November 2023. DeepSeek-V2 came in May 2024 and DeepSeek-V3 in December 2024.

DeepSeek-R1 arrived on 20 January 2025. DeepSeek open-sourced six smaller distilled models with it, including 32B and 70B versions that its release note puts on par with OpenAI-o1-mini. Wikipedia adds that the distilled models were initialized from other models, such as Llama and Qwen, and trained on data synthesized by R1 and R1-Zero.

DeepSeek-V3.1, in August 2025, brought a hybrid design with thinking and non-thinking modes. DeepSeek-V3.2 came in December 2025. Wikipedia lists the newest generation, DeepSeek-V4, in April 2026, with V4-Pro and V4-Flash models and a one million token context window. The DeepSeek-V4-Pro model card gives it 1.6 trillion parameters, 49 billion of them activated.

Licensing pattern

DeepSeek-V3 split its licensing in two. The repository code is MIT licensed. The weights sit under a separate Model License, which the model card says supports commercial use.

DeepSeek-R1 changed that. Its release note says it is MIT licensed and spells out that users can use the weights, and the outputs of its API, for fine-tuning and distillation. Wikipedia says DeepSeek has released its models under free and open-source software licenses since January 2025, and both Wikipedia and the DeepSeek-V4-Pro card list the MIT License for V4.

The training data stays closed. DeepSeek shares the exact parameters of its models, but per Wikipedia the training data is not openly licensed.

Release lineage

Newest first
  1. DeepSeek-V4.1-Flash-1017763B
  2. DeepSeek-V4.1-Flash763B
  3. DeepSeek-V4-Flash-Vision-Exp305B
  4. DeepSeek-V4-Pro-08131.65T
  5. DeepSeek-V4-Flash-0731304B
  6. DeepSeek-V4-Flash-DSpark304B
  7. DeepSeek-V4-Pro1.6T
  8. DeepSeek-V4-Flash291B
  9. DeepSeek-OCR-23.4B
  10. DeepSeek-V3.2685B
  11. DeepSeek-OCR3.3B
  12. DeepSeek-R1-0528-Qwen3-8B8.2B

Largest first. The last line on each tile is the memory of the smallest device that fits the model at 4-bit.