Chinese start-up DeepSeek on Tuesday released version 3.1 of its flagship large-language model, expanding the maximum context window to 128,000 tokens—roughly four times the length of a typical novel. The update, announced through the company’s official WeChat channel, is live on DeepSeek’s website, mobile app and mini-program, and can be accessed through the existing API without code changes. A base version of V3.1 has also been uploaded to the open-source hub Hugging Face, while third-party platform Chutes lists pay-as-you-go rates of $0.1999 per million input tokens and $0.8001 per million output tokens. The larger context window allows the model to ingest and recall substantially more text in a single interaction, a feature prized by enterprise users building document-heavy applications. DeepSeek’s latest incremental release keeps attention on the Hangzhou-based company as it races U.S. rivals such as OpenAI to push model capacity at lower cost. The upgrade arrives amid repeated delays to DeepSeek’s next-generation R2 system, which local media attribute to chief executive Liang Wenfeng’s insistence on further tuning. Even so, the V-series’ rapid cadence has helped the six-month-old firm carve out a sizable domestic user base and intensify global competition in generative AI.