DeepSeek V3 has released its latest version, V3-0324. Although the company refers to it as a "minor update," practical tests reveal that the performance improvements are quite significant, even rivaling the capabilities of the V3.5 version.

In testing the new version, DeepSeek V3-0324 demonstrated remarkable capabilities. In a complex bouncing ball test, the model successfully tackled the challenge of a 4-dimensional hypercube, showcasing its proficiency in handling high-dimensional spatial problems.

When it comes to programming, DeepSeek V3-0324 also excels. With just a single prompt, it can generate a complete product landing page with adaptive layout and animations. This functionality is comparable to Claude 3.7 Sonnet, highlighting the robust content generation abilities of this new version.

In developer Xeophon's personal benchmark tests, DeepSeek V3-0324 achieved notable improvements across all metrics, emerging as the top-performing non-reasoning model in these evaluations.

Notably, while DeepSeek V3-0324 is not a reasoning model, it still exhibits a degree of cognitive decomposition when addressing complex problems. When facing difficult questions, the model can autonomously revisit previous steps for reconsideration and display an "aha moment" by identifying hidden conditions not explicitly mentioned in the problem.

Additionally, DeepSeek V3-0324 remains free and open-source, with its weight files now available on HuggingFace under the most permissive MIT license. The total disk space required for all weight files is approximately 688GB, consistent with the initial V3 version, indicating it is still a 671-billion-parameter Mixture-of-Experts (MoE) model.

Currently, users can access DeepSeek V3-0324 through the official website, the official app (with deep thinking mode disabled), and platforms like HuggingFace. Furthermore, the model has been entered into the large model arena to compete with other models, and the voting results will be announced in the near future.