DeepSeek releases DeepSeek-V3-0324 update
The updated checkpoint scored 81.2% on MMLU-Pro and 59.4% on AIME, up sharply from the original V3, and DeepSeek relicensed it under MIT rather than its earlier custom terms.
- Open weights & ecosystem
- Models & capabilities
- Minor
DeepSeek released an updated checkpoint of its V3 model, DeepSeek-V3-0324, without changing the underlying architecture or parameter count. The company reported gains across its benchmark suite: MMLU-Pro rose from 75.9% to 81.2%, GPQA from 59.1% to 68.4%, AIME from 39.6% to 59.4%, and LiveCodeBench from 39.2% to 49.2%, alongside qualitative claims of stronger front-end coding and Chinese-language writing.
The more consequential change was licensing: DeepSeek moved the weights from V3’s earlier custom licence to the permissive MIT licence, matching the terms it had already applied to its reasoning model R1 two months earlier. That removed restrictions on commercial use and redistribution, letting the update flow directly into the ecosystem of hosted providers and downstream fine-tunes that had built up around DeepSeek’s models since R1’s January 2025 release.
The update illustrated a pattern that had become familiar with DeepSeek: substantial capability gains delivered as an incremental checkpoint refresh rather than a headline new model, published quietly via API documentation and Hugging Face rather than a dedicated announcement. Coming roughly two months after R1 had triggered a market reaction over the cost of Chinese open-weight training, the V3-0324 update reinforced that DeepSeek’s releases were continuing to narrow the reported gap with closed frontier models on reasoning and coding benchmarks.