DeepSeek

DeepSeek · 7 tracked releases · May 6, 2024 – September 10, 2026

At a glance

Releases tracked7
First trackedMay 6, 2024 — DeepSeek-V2
LatestSeptember 10, 2026 — DeepSeek-V4.1-Flash
Span857 days
Median gap139 days between releases
WeightsOpen throughout
Record quality6 of 7 verified · how this is judged

Release cadence

One tile per tracked release, placed in the month it shipped, so the pauses in this line are the pauses in the data. Years with nothing tracked keep their row.

  1. 2024
    1. DeepSeek-V2, DeepSeek, May 6, 2024
    2. DeepSeek-V2.5, DeepSeek, September 5, 2024
    3. DeepSeek-V3, DeepSeek, December 26, 2024
    3
  2. 2025
    1. DeepSeek-R1, DeepSeek, January 20, 2025
    2. DeepSeek-V3.1, DeepSeek, August 21, 2025
    2
  3. 2026
    1. DeepSeek V4 Flash, DeepSeek, April 24, 2026
    2. DeepSeek-V4.1-Flash, DeepSeek, September 10, 2026
    2

Lineage

Ordered by announcement date, and spaced by the gap between releases, so a quiet year looks like one. Lineage is not inferred from version numbers — labs do not number consistently, so the dates decide the order.

  1. DeepSeek-V2Text in and out

    coding

  2. +122 days

    DeepSeek-V2.5Text in and out

    function callingstructured output

    from DeepSeek-V2 first evidenced: function calling, structured output

  3. +112 days

    DeepSeek-V3Text in and out

    from DeepSeek-V2.5 Parameters 236B → 671B

  4. +25 days

    DeepSeek-R1Text in and out

    reasoning

  5. +213 days

    DeepSeek-V3.1Text in and out

    agentictool use

    from DeepSeek-V3 Licence DeepSeek License Agreement 1.0 → MITfirst evidenced: agentic, function calling, reasoning, tool use

  6. +246 days

    DeepSeek V4 FlashText in and out
  7. +139 days

    DeepSeek-V4.1-FlashText in and outImage input

    from DeepSeek V4 Flash first evidenced: image in

Context window over time

Only releases with a disclosed context window appear — 2 of 7 are not shown.

DeepSeek-V2 · 2024 128K tokens
DeepSeek-V3 · 2024 128K tokens
DeepSeek-R1 · 2025 128K tokens
DeepSeek-V3.1 · 2025 128K tokens

What changed

Generation to generation, from evidenced values only. A field is compared only where both releases record it — otherwise it is called out as not comparable, so a research gap never reads as a change the lab made.

DeepSeek-V2 → DeepSeek-V2.5 122 days

Capabilities first evidenced
+ function calling + structured output

Not comparable here: context window (not disclosed for both releases).

DeepSeek-V2.5 → DeepSeek-V3 112 days

Parameters
236B → 671B

Not comparable here: context window (not disclosed for both releases).

DeepSeek-V3 → DeepSeek-V3.1 238 days

Licence
DeepSeek License Agreement 1.0 → MIT
Capabilities first evidenced
+ agentic + function calling + reasoning + tool use

DeepSeek V4 Flash → DeepSeek-V4.1-Flash 139 days

Modalities first evidenced
+ image in

Not comparable here: context window (not disclosed for both releases); parameters (not disclosed for both releases).

Compare the two most recent side by side →
All releases from DeepSeek →