DeepSeek V3-0324

2025-03-24 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases V3-0324, an update to its 685-billion-parameter generalist model, with a clear gain in reasoning, web development and tool calling, now under the MIT license.

### Context V3 (December 2024) had established DeepSeek as a leading open lab, but under a model-specific license. R1, released two months earlier, had already been placed under the MIT license. V3-0324 aligns the generalist model with that license and fixes several flaws of the previous version, in a market where GPT-4.5 and Claude 3.7 Sonnet were the closed models the vendor chose for the comparison. ### What the model brings - A marked gain in reasoning without a “thinking” mode: according to the model card, +5.3 points on MMLU-Pro, +9.3 on GPQA and +19.8 on AIME compared with V3. - Better front-end development skills, with web code that runs more often and more polished pages. - Stronger Chinese writing, and better use of search based on web results. - More accurate function calling, with the previous version's errors fixed. - An MIT license, like R1, for the code and the weights. ### Architecture and specifications - Type: Mixture-of-Experts, 685 billion parameters in total (value shown by Hugging Face). The vendor states that the structure is identical to that of V3 (671 billion parameters for the main model, plus additional modules). Active parameters, number of experts and context: not included in the card consulted. - Modalities: text. Training data and knowledge cutoff: not disclosed. - License: MIT (code and weights), commercial use allowed, with no threshold. - Recommended setting: temperature 0.3, with a formula mapping the API temperature to the model's temperature. - Weights: huggingface.co/deepseek-ai/DeepSeek-V3-0324. ### Cost and access | Item | Detail | |---|---| | API usage | Unchanged, identifier `deepseek-chat` | | Pricing at launch | Not disclosed in the post | | Self-hosting hardware | Not disclosed | ### Benchmarks | Benchmark | DeepSeek-V3 | DeepSeek-V3-0324 | Qwen-Max | GPT-4.5 | Claude-Sonnet-3.7 | |---|---|---|---|---|---| | MMLU-Pro (EM) | 75.9 | 81.2 | 76.1 | 86.1 | 80.7 | | GPQA Diamond (Pass@1) | 59.1 | 68.4 | 60.1 | 71.4 | 68.0 | | MATH-500 (Pass@1) | 90.2 | 94.0 | 82.6 | 90.7 | 82.2 | | AIME 2024 (Pass@1) | 39.6 | 59.4 | 26.7 | 36.7 | 23.3 | | LiveCodeBench (Pass@1) | 39.2 | 49.2 | 38.7 | 44.4 | 42.2 | Source: chart in the vendor's official post, scores published by the vendor. ### Versus closed models At the time of its release, the vendor compares V3-0324 with GPT-4.5 and Claude Sonnet 3.7, two closed models. According to its chart, V3-0324 outperforms GPT-4.5 on MATH-500 (94.0 versus 90.7), AIME 2024 (59.4 versus 36.7) and LiveCodeBench (49.2 versus 44.4). It remains behind GPT-4.5 on MMLU-Pro (81.2 versus 86.1) and GPQA Diamond (68.4 versus 71.4). Against Claude Sonnet 3.7, it is ahead on MMLU-Pro (81.2 versus 80.7) and on a par on GPQA (68.4 versus 68.0). The post gives no pricing, so the cost comparison is not quantified. For the enterprise, the main benefit lies in the MIT license, which removes any ambiguity about commercial use and hosting in one's own environment. ### Key takeaways V3-0324 shows that an open generalist model can catch up with closed models of the same period on several published benchmarks. The move to the MIT license makes it a simple choice to contract for.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Code, Appel de fonctions