DeepSeek-V3-0324

What it is

DeepSeek-V3-0324 is an updated checkpoint of DeepSeek-V3, the large language model from DeepSeek. On Hugging Face the model is listed at 685 billion parameters, and its model card states that the model structure is exactly the same as DeepSeek-V3's ^[2]. A checkpoint is a saved copy of a model's trained weights, so an update of this kind changes what the model has learned without changing the shape of the network or the way it is served. The weights themselves were published on Hugging Face ^[1].

Release and licence

Wikipedia dates the update to 24 March 2025, while DeepSeek's own release note is dated 25 March 2025 ^[4][1]. The licence changed with this release: the model came out under the MIT License, the same licence used for the reasoning model DeepSeek-R1, replacing the licence the original V3 model had been released under ^[1]. MIT is a permissive open-source licence that allows modification and redistribution.

What changed

DeepSeek cited a major boost in reasoning performance, stronger front-end web development skills and smarter tool use ^[1]. Nothing changed on the API side: API usage stayed the same, and DeepSeek recommended running the model with DeepThink turned off for tasks that do not require complex reasoning ^[1]. DeepThink is the switch that makes a model write out a long chain of reasoning before answering; leaving it off keeps responses shorter for simple work.

Benchmark results

Two sets of numbers describe the update. The Hugging Face model card compares the new checkpoint with the original V3 on four benchmarks: MMLU-Pro goes 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 ^[2]. MMLU-Pro is a broad multiple-choice test of knowledge, GPQA a set of graduate-level science questions, AIME a competition mathematics exam, and LiveCodeBench a coding benchmark drawn from programming contests.

DeepSeek's later comparison table gives 81.2 on MMLU-Pro, 68.4 on GPQA-Diamond and 51.3 on AIME 2025 ^[3]. The same table records 45.4 on SWE-bench Verified in agent mode, 13.3 on Terminal-bench and 43.0 on LiveCodeBench (2408-2505) ^[3]. SWE-bench Verified asks whether a model can fix real issues in software repositories, Terminal-bench tests work in a command-line environment, and the numbers in brackets on the LiveCodeBench entry mark the date range of the problems used. The two sources report different LiveCodeBench scores: 49.2 in the model card, 43.0 in the later table ^[2][3].

Where it sits

DeepSeek-V3-0324 was succeeded by DeepSeek-V3.1 on 21 August 2025 ^[3]. In that model's comparison table, V3.1's non-thinking mode, which answers directly instead of producing a long reasoning trace, scores 83.7 on MMLU-Pro and 66.0 on SWE-bench Verified ^[3]. Because the 0324 update shares its structure with the original V3, it reads as one step in a series of releases built on the same base model rather than as a separate model family.