What happened
DeepSeek released an updated DeepSeek-V3 checkpoint, DeepSeek-V3-0324. Wikipedia dates the update to 24 March 2025, and DeepSeek's release note is dated 25 March 2025. [4][1]
The update replaced the original V3 model license with the MIT License, the same license DeepSeek-R1 uses. [1]
A checkpoint is a saved set of trained model weights. An updated checkpoint replaces an earlier one within the same model family, so release notes are the main place to see what changed.
Key details
DeepSeek cited a major boost in reasoning performance, stronger front-end web development skills and smarter tool use. [1]
API usage was unchanged. DeepSeek recommended running the model with DeepThink turned off for tasks that are not complex reasoning. [1]
On a later DeepSeek comparison table the model scored 81.2 on MMLU-Pro, 68.4 on GPQA-Diamond and 51.3 on AIME 2025. [3]
The weights were published on Hugging Face. [1][2]
Impact
The change of license is the most concrete part of the update for downstream users. The MIT License is one of the most permissive open-source licenses, mainly requiring that the copyright notice be kept, so people building on the weights face few conditions. The benchmark figures are DeepSeek's own comparisons. Standardized test sets such as MMLU-Pro, GPQA-Diamond and AIME 2025 only give fair comparisons when results are measured the same way.
Sources
- 1DeepSeek-V3-0324 ReleaseDeepSeek API Docs
- 2deepseek-ai/DeepSeek-V3-0324Hugging Face
- 3deepseek-ai/DeepSeek-V3.1Hugging Face
- 4DeepSeekWikipedia
Cite this page: DeepSeek updates V3 as DeepSeek-V3-0324 under the MIT License. TellWorld, https://www.tellworld.net/news/deepseek-v3-0324-2025. Updated 2026-11-08