Yesterday, DeepSeek officially released the technical update report for its latest model, DeepSeek-V3. The new version has demonstrated superior performance in math and coding-related benchmarks, even surpassing OpenAI’s GPT-4.5.
DeepSeek-V3 model demonstrates superior performance
Notably, these improvements were achieved solely through enhancements in training methods. DeepSeek-V3-0324 is based on the same foundational model as the previous DeepSeek-V3, debunking earlier rumors that suggested it was built on the R2 base model.
Key Updates in DeepSeek-V3
The latest model features approximately 660 billion parameters, slightly different from the previously rumored 685B. The open-source version supports a 128K context length, while the web, app, and API versions provide a 64K context.
For private deployment**, users only need to update the checkpoint and tokenizer_config.json files, with adjustments related to tool calls.
To experience this updated model, users can simply log in to the official website, app, or mini-program and disable the “Deep Thinking” feature in the conversation interface. The API interface and usage remain unchanged.
DeepSeek recommends using the V3 model for non-complex reasoning tasks, as it offers enhanced performance.
Performance Enhancements Across Multiple Domains
- Frontend Development: The generated code is now more functional and visually optimized.
- Chinese Writing: Significant improvements in medium-to-long-form content quality compared to the R1 version.
- Chinese Search: With internet access enabled, search results are now more detailed, accurate, and better formatted.
- Other Areas: Upgrades in tool usage, role-playing, Q&A, and casual conversations.
Open-Source Flexibility and Deployment
The model is released under the permissive MIT open-source license and can be directly deployed on an M3 Ultra Mac Studio. This significantly lowers the barrier for developers looking to build large-model applications.
Many users have already tested the model’s capabilities, including interactive tasks like simple game development. With its latest advancements, DeepSeek-V3 is setting a new benchmark in AI performance and accessibility.
