Chinese company DeepSeek has released update V3-0324 for its artificial intelligence model. The startup says the update improves the AI’s programming and mathematical problem-solving capabilities, and increases the number of model parameters to 685 billion.
Chinese company DeepSeek has released update V3-0324 for its artificial intelligence model. The startup says the update improves the AI’s programming and mathematical problem-solving capabilities, and increases the number of model parameters to 685 billion.
DeepSeek-V3-0324, named after its predecessor and launch date, has «enhanced reasoning capabilities, optimized front-end web development, and improved Chinese writing,» the South China Morning Post reported.
The updated fundamental model improved results in several benchmarks, most notably in the American Mathematics Exam (AIME), where it scored 59.4 points compared to 39.6 in the previous version, and on LiveCodeBench it increased its score by 10 points to 49.2 points, DeepSeek notes.
Compared to DeepSeek V3, which has 671 billion parameters and uses the company’s own commercial license, the new model with 685 billion parameters uses the MIT software license, which is the most popular on the GitHub developer platform.
Launched on the Hugging Face AI community as well as the company’s own website, DeepSeek-V3-0324 is currently the most popular model on Hugging Face, receiving positive feedback for its performance.
Jasper Zhang, a gold medalist in the Mathematical Olympiad who graduated from the University of California, Berkeley with a doctorate, tested the model on the AIME 2025 problem, and «it solved it without a problem.»
«More confident open-source AI models will ultimately win», — Zhang said on X (Twitter). He added that his startup Hyperbolic now supports DeepSeek-V3-0324 on its cloud platform.
Recall that at the end of December 2024, it became known that the Chinese company DeepSeek introduced its new open AI model — DeepSeek V3, which resembles ChatGPT.
A month later, on January 20, the company introduced a new version of its AI, DeepSeek-R1. The developers claim that it is not inferior to OpenAI’s «thoughtful» o1 model in terms of performance and affordability.
Just a week after the presentation of the new model, shares of Asian tech companies began to fall.
Since R1 was released a few weeks after DeepSeek-V3, there is speculation that a new reasoning model could be introduced shortly after DeepSeek-V3-0324. DeepSeek had planned to release R2 in early May, but may do so earlier, Reuters reported.
«The programming capabilities are much stronger, and the new version could pave the way for the launch of R2», — said Li Bangzhu, founder of AIcpb.com, a website that tracks the popularity of artificial intelligence applications.
