DeepSeek-V3-0324: the quiet V3 update, now MIT licensed
In one sentence DeepSeek publishes an update of DeepSeek-V3 on Hugging Face (685B parameters, open weights under the MIT license). Big reasoning gains: AIME from 39.6 to 59.4, GPQA from 59.1 to 68.4. On the Aider Polyglot benchmark it scores 55.1%, at a far lower cost than closed competitors.
On March 24, 2025 DeepSeek uploads to Hugging Face, with almost no announcement, a new version of its flagship model: DeepSeek-V3-0324. The official announcement follows the next day.
It is not a new model from scratch but an update of V3 (released December 2024), with the same Mixture of Experts architecture: many internal "experts", only some of which activate for each word. The license changes: weights and code are now under MIT, so they can be downloaded and used commercially.
The gains show mostly in reasoning and coding: on competition math problems (AIME) it goes from 39.6% to 59.4%. On Aider's coding benchmark it solves 55.1% of the exercises: among models without a reasoning phase only Claude 3.7 Sonnet (60.4%) does better, and the whole test costs about one dollar. API usage stays unchanged.
Companies
DeepSeek
Tools
DeepSeek-V3-0324
Tags
Sources