DeepSeek has shipped two main model lines since late 2024: the V-series general models and the R-series reasoning models, with the two merging into a single architecture from V3.1 onward. Here is every release DeepSeek has announced, in order, with what changed and whether the weights were published.

On this page

2024–2025: V3, R1 and their updates

DeepSeek released DeepSeek-V3 on December 26, 2024, publishing the weights on Hugging Face alongside a technical report. V3’s weights shipped under DeepSeek’s own custom model license at the time, separate from the MIT-licensed training and inference code.

DeepSeek-R1, a reasoning model trained from the V3 base, followed on January 20, 2025. DeepSeek switched the licensing approach for this release, putting the R1 weights themselves under the MIT license rather than its earlier custom terms — the release that drew most of the outside attention to the company.

DeepSeek updated the general-purpose line with DeepSeek-V3-0324 on March 24–25, 2025: a re-trained version of V3 that DeepSeek said improved coding and front-end generation and made the model more willing to use tools. DeepSeek-R1-0528 followed on May 28, 2025, a smaller update to R1 rather than a new generation.

DeepSeek then began merging the two lines. DeepSeek-V3.1, released in August 2025, added a hybrid mode: the same weights can run in a fast “non-thinking” mode or switch into R1-style step-by-step reasoning on request, removing the need for separate chat and reasoning models. DeepSeek-V3.1-Terminus followed on September 22, 2025 as an incremental update to that hybrid model.

DeepSeek closed out the V3 generation with DeepSeek-V3.2 on December 1, 2025, released with a reasoning-focused V3.2-Speciale variant alongside it. Both shipped as open weights under the MIT license.

2026: the V4 generation

DeepSeek opened its next architecture generation with a DeepSeek-V4 preview on April 24, 2026, giving outside developers open-weight access to the new line along with a 1-million-token context window and upgraded agentic and coding behavior.

The V4 line then split into a flagship and a fast tier, each updated separately:

  • DeepSeek-V4-Flash reached public beta on July 31, 2026, as a re-trained version of the earlier preview build, adapted for coding tools and the Responses API format.
  • DeepSeek-V4-Pro reached general availability on August 13, 2026, across DeepSeek’s app, web interface and API, adding native support for the Responses API format and three selectable thinking-effort levels (low, high, max).
  • DeepSeek-V4-Flash-Vision-Exp, an experimental vision-capable variant, followed on August 21, 2026.
  • DeepSeek retired both Flash variants on September 10, 2026 and replaced them the same day with DeepSeek-V4.1-Flash, the smallest model in the V4 family and the first with native multimodal image understanding; the old model names now route to it automatically.
  • DeepSeek updated that model again on October 17, 2026 as DeepSeek-V4.1-Flash-1017, with the same API name.

As of this writing, DeepSeek-V4-Pro and DeepSeek-V4.1-Flash are the current, generally available models in the V4 line; every earlier V4-generation name listed above has been superseded or retired.

What’s open-weight and what isn’t

DeepSeek has published open weights for most of its major releases. V3 shipped under DeepSeek’s own custom model license; starting with R1 on January 20, 2025, DeepSeek switched to the MIT license, and R1-0528, V3.1, V3.2 and the V4 preview all followed under those same MIT terms. The later V4-generation production updates (V4-Pro, V4.1-Flash and its October update) have so far been documented by DeepSeek only as API, app and web releases rather than weight drops. For the exact terms of any weights release, read the licence file published with it by DeepSeek.