DeepSeek-V3-0324 launched on the AI development platform Hugging Face

DeepSeek from China has launched an upgrade to its V3 large language model, DeepSeek-V3-0324, on the AI development platform Hugging Face. The startup marketed as including improvements in reasoning and coding capabilities over its earlier V3 model.

DeepSeek-V3-0324 demonstrates notable improvements over its predecessor, DeepSeek-V3, in several key aspects. There are significant improvements in benchmark performance, such as reasoning capability, front-end Web development, where it improved executability of the code. There are more aesthetically pleasing web pages and game front-ends. Also, Chinese writing proficiency has significantly increased, with enhanced style and content quality.

Chinese search capabilities has enhanced report analysis requests with more detailed outputs. There are function calling improvements too, with an increased accuracy in Function Calling, fixing issues from previous V3 versions.

The model structure of DeepSeek-V3-0324 is exactly the same as DeepSeek-V3. This model supports features such as function calling, JSON output, and FIM completion. Do note that Hugging Face's Transformers has not been directly supported yet.

DeepSeek-V3-0324 has outperformed DeepSeek-V3, Qwen Max, GPT 4.5, and Claude-Sonnet-3.7 in all areas, as per the figure above. For non-complex reasoning tasks, we recommend using V3 — just turn off “DeepThink”. The API usage remains unchanged. Models are now released under the MIT License, just like DeepSeek-R1!

As per Reddit, the DeepSeek V3-0324 marks the first time an open weights model has been the leading non-reasoning model. This is the third release from the company, who launched the V3 model in Dec. 2024, followed by release of the R1 model in Jan. 2025.