Chinese AI company DeepSeek open-sourced its new large language model, V3.1-Base, on the Hugging Face platform late last night, sparking renewed interest from the developer community. According to official information, the model boasts 685 billion parameters, with a key highlight being an expanded context window of 128KB. This means the model can simultaneously process extremely long texts, equivalent to 300,000 Chinese characters, significantly improving its ability to handle complex tasks. The open-source link has also been published on the Hugging Face repository, allowing developers to freely download and test it.
Notably, just hours before the open-source release, DeepSeek completed a comprehensive upgrade of its online services. Users can now directly experience V3.1 through its official website, mobile app, and mini-program. API calls remain compatible, eliminating the need for enterprise users to adjust their existing integration solutions. This simultaneous update demonstrates DeepSeek's coordinated approach to open source and commercialization, and its rapid technological iteration has attracted industry attention.
However, the highly anticipated DeepSeek R2 project has yet to announce a specific release schedule. Analysts point out that the current V3.1 version's performance breakthroughs in scenarios like long text understanding and multi-round conversations may lay an important technical foundation for subsequent product lines. With 128KB of context becoming the new industry benchmark, the large-scale model competition led by domestic teams is entering a new phase.
