DeepSeek V3-0324: A Checkpoint Update That Moved AIME by 20 Points

What V3-0324 is

DeepSeek-V3-0324 is a checkpoint update to DeepSeek V3 rather than a new architecture, and the gains it reports are larger than that framing suggests. DeepSeek publish the before-and-after directly: MMLU-Pro rose from 75.9 to 81.2, GPQA from 59.1 to 68.4, LiveCodeBench from 39.2 to 49.2, and AIME from 39.6 to 59.4 — a twenty-point jump on competition mathematics from what is nominally a point release. It is open-weight and self-hostable, as the whole V3 line is.

The temperature mapping

The card's usage notes describe a temperature mapping. DeepSeek run their own web and app products at a model temperature of 0.3. Because most API callers leave temperature at the default 1.0, DeepSeek implemented a mapping that silently rescales it: an API temperature between 0 and 1 is multiplied by 0.3, so calling with 1.0 gives you a model temperature of 0.3. Above 1.0 the mapping changes to subtracting 0.7. The consequence is that the temperature you set is not the temperature you get, and anyone porting prompts between DeepSeek and another provider — or comparing outputs across them — is comparing different sampling regimes without being told.

What else changed, in DeepSeek's words

Beyond the benchmark movement, DeepSeek list improvements that are mostly about usability rather than capability. Front-end web development produces more executable code and better-looking pages and game front-ends. Chinese writing was aligned with the R1 writing style, with better medium-to-long-form quality, improved multi-turn interactive rewriting, and better translation and letter writing. Chinese search produces more detailed report analysis. And function calling accuracy was increased, explicitly fixing issues present in earlier V3 versions — worth knowing if you evaluated the original V3 for tool use and rejected it, because that specific complaint was addressed here. DeepSeek also publish the system prompt their own products use, which names the assistant and supplies the current date.

LMArena and context window

On LMArena's text leaderboard DeepSeek V3 0324 has 1395 Elo (rank 169); by category, 1428 Elo in Coding (rank 190), 1367 Elo in Math (rank 203), 1397 Elo in Expert (rank 186) and 1390 Elo in Creative Writing (rank 130). It carries a 32K context window; later DeepSeek releases reach 128K and the V4 generation reaches 1 million. It is around not published per million input tokens through third-party providers.

Features, prompts and licence

The model structure is exactly the same as DeepSeek-V3; DeepSeek point to the V3 repository for running it locally, and note that Hugging Face Transformers did not support it directly at release. It supports function calling, JSON output and fill-in-the-middle completion, with prompt formats documented in the DeepSeek-V2.5 repository. DeepSeek publish separate prompt templates for file upload, with file name, file content and question as arguments, and for web search, with search results, the current date and the question; the search template instructs the model to cite results inline in the form [citation:X] and to limit list answers to ten points. The repository and weights are licensed under MIT.

Official Resources