What's New in DeepSeek's Latest V3-0324 Update? 177

小编 ·

What's New in DeepSeek's Latest V3-0324 Update?

DeepSeek V3 received a minor version upgrade on March 24, 2025 (build V3-0324), which has also been pushed to the official website. To use it, simply turn off deep thinking mode. Here's a breakdown of the changes:

What's New in DeepSeek's Latest V3-0324 Update?

DeepSeek V3 received a minor version upgrade on March 24, 2025, with the version number V3-0324, and it has also been updated synchronously on the official website. To use it, simply turn off deep thinking. The update details are as follows:

What updates does the latest version of DeepSeek V3-0324 include?.jpg

Performance Improvements

  • Coding capability significantly enhanced: The new version of V3 shows notable progress in front-end coding capabilities, with substantial improvements in code quality and logical clarity. Particularly in HTML, CSS, and JavaScript, it's no exaggeration to say that the V3-0324 version has thoroughly surpassed its previous self, and its coding capability is now quite close to Claude's.

  • Enhanced math capabilities: The new version of V3 supports complex problem-solving, reaching the level of a reasoning model, with a significant improvement in math capabilities.

  • Improved multi-turn conversation capabilities: The new version of V3 demonstrates stronger context understanding in multi-turn conversations, with the context length doubling from 64K to 128K, supporting longer conversation history and processing of lengthy documents.

  • Improved generation speed: DeepSeek-V3's generation speed has increased from 20 TPS to 60 TPS, achieving a 3x improvement compared to the V2.5 model.

Open Source License Change

  • The new version of V3 adopts the MIT open source license. Compared to the initial V3 open source license, this change allows developers to use the model more freely in commercial projects, greatly enhancing its flexibility and appeal in commercial applications.

Model Parameters and Training

  • The open source version of the new V3 model has a size of 685 billion parameters, supported by improved post-training on the new 32K GPU cluster.

User Experience

  • Real-world test feedback: In a task of generating an animated weather card containing CSS and JavaScript, the result generated by the new version of V3 was not only logically clear but also excellent in effect, significantly outperforming the older version R1.

  • Comparison with other models: Comparison tests with Claude 3.7 show that the new version of V3 is already comparable to it in front-end website building.

Other Updates

  • API interface unchanged: Users can call the new version of V3 by specifying model=deepseek-chat.

  • Trial experience: Users can experience this new version by visiting DeepSeek's official website, application, and mini-program.

HuggingFace:https://huggingface.co/deepseek-ai/DeepSeek-V3-0324

Project address:https://github.com/deepseek-ai/DeepSeek-V3