한국어EN
DeepSeek

The DeepSeek family

Models from China's DeepSeek, mostly released as open weights under permissive licenses such as MIT.

Only versions we have a record for. Missing ones appear as records are added.

12versions2023–2026Latest · DeepSeek-V4.1-Flash

At a glance

The DeepSeek family — 12 versionsEach DeepSeek version from 2023 to 2026 as a dot, in release order. The list below has the same information.Main lineDeepSeek LLM · Nov 29, 2023DeepSeek-V2 · May 7, 2024DeepSeek-V3 · Dec 2024DeepSeek-R1 · Jan 20, 2025DeepSeek-V3.1 · Aug 21, 2025DeepSeek-V3.2-Exp · Sep 29, 2025DeepSeek-V3.2 · Dec 1, 2025DeepSeek V4-Pro / V4-Flash · Apr 24, 2026DeepSeek-V4.1-Flash · Sep 17, 2026Specialist modelsDeepSeekMath · Feb 5, 2024DeepSeek-Coder-V2 · Jun 17, 2024DeepSeek-OCR · Oct 21, 2025
Where labels would overlap, only the dot is drawn — point at it to see the name; every version is in the list below.

12 versions

In release order. Each date comes from the record that documents it — open a name to see the record.

  1. DeepSeek LLM

    Released its first general models, 7B and 67B

    Main lineTextOpen weights

  2. DeepSeekMath

    Paper that introduced the GRPO training method

    Specialist modelsText

  3. DeepSeek-V2

    Introduced MLA, which compresses the KV cache

    Main lineOpen weights

  4. DeepSeek-Coder-V2

    A code model trained on from DeepSeek-V2

    Specialist modelsCodeOpen weights

  5. DeepSeek-V3

    MIT-licensed MoE noted for its cost efficiency

    Main lineOpen weights

  6. DeepSeek-R1

    Open reasoning model under MIT, same day as Kimi K1.5

    Main lineOpen weightsReasoning

  7. DeepSeek-V3.1

    One model that switches between thinking and non-thinking modes

    Main lineReasoningOpen weights

  8. DeepSeek-V3.2-Exp

    Experimental release debuting DeepSeek Sparse Attention

    Main lineText

  9. DeepSeek-OCR

    Uses images as a way to compress context

    Specialist modelsMultimodal

  10. DeepSeek-V3.2

    Official successor to V3.2-Exp, released with V3.2-Speciale

    Main lineReasoningOpen weights

  11. DeepSeek V4-Pro / V4-Flash

    MIT-licensed, with a 1M-token context

    Main lineOpen weights

  12. DeepSeek-V4.1-Flash

    Multimodal MoE built to shrink the KV cache

    LatestMain lineMultimodal