DeepSeek releases V4 as open weights under MIT
Originally announced by DeepSeek . We link the primary source so you can read it for yourself.
DeepSeek has released DeepSeek V4, reported to comprise two models, V4-Pro and V4-Flash, under the permissive MIT licence and available as open weights. Coming after the R1 reasoning model and the V3 line, it continues the lab’s pattern of shipping capable open-weight models under unusually permissive terms.
What was released
According to reporting around the launch, V4 arrived as two production models rather than a preview, with a long context window and a reworked approach to attention aimed at handling long inputs more efficiently. As ever, the specifics and the performance figures are the developer’s own or early reporting, so they warrant the same caution we apply to any results that have not been independently confirmed.
Why it matters
Attention is the part of a transformer whose cost grows sharply with context length, so genuine efficiency gains there matter a great deal for running long-context workloads locally. Equally important for this audience is the licence: an MIT release means the models can be used, modified, and self-hosted commercially without conditions, as with R1 before them.
The context
DeepSeek has been one of the more consequential open-weight labs, in part because R1 was released under MIT at a capability level that surprised many. A V4 generation on the same permissive footing continues a clear 2026 trend of strong open-weight models shipping under genuinely open licences.
We will assess V4 for the catalogue once independent evaluations and its practical hardware requirements are available.