DeepSeek announced the release of an upgraded version of its V3 model on Tuesday, which includes improvements in areas like front-end development, Chinese writing and tool-use capabilities that are not reliant on complex multi-step reasoning.

[Photo/Deepseek]

According to the latest benchmark data released by DeepSeek, V3-0324 outperforms the original V3 in multiple categories. On the MMLU-Pro benchmark, which tests general knowledge and reasoning, accuracy rose from 75.9 percent to 81.2 percent. The GPQA Diamond benchmark, which evaluates models on graduate-level questions across a range of scientific disciplines, saw a jump from 59.1 percent to 68.4 percent. Coding proficiency of the newer model has also increased, with LiveCodeBench results moving from 39.2 percent to 49.2 percent. These benchmark results reflect the new model’s precision and reliability across a range of academic and practical tasks.

The update also strengthens the model’s front-end development abilities, with more aesthetically refined and executable outputs for both websites and game interfaces. In Chinese writing, the model now generates content that aligns more closely with the DeepSeek R1 writing style, with noticeable improvements in medium to long-form content quality.

Some other areas of improvement include better performance in interactive rewriting, enhanced translation and letter-writing features, and more detailed responses to Chinese-language search queries. 

For users engaging in non-complex reasoning tasks, DeepSeek recommends using V3 with the “DeepThink” mode turned off, maintaining efficiency without sacrificing output quality.