On the evening of September 22, DeepSeek rolled out its latest upgraded model, DeepSeek-V3.1-Terminus, on its official API platform and simultaneously made it open-source. Leveraging its existing strengths, this new iteration has enhanced language consistency, resolved occasional issues with abnormal characters, and fine-tuned the performance of its programming and search functionalities. The DeepSeek-V3.1-Terminus model comes equipped with both thinking and non-thinking modes, each capable of handling a context length of up to 128K, thereby catering to the demands of intricate long-text tasks. In the non-thinking mode, the default output length is set at 4K, with the option to extend up to 8K. Conversely, in the thinking mode, the default length is 32K, with a maximum reachable length of 64K. When it comes to pricing, the model's input costs vary depending on cache utilization: a mere 0.5 yuan per million tokens for cache hits, and 4 yuan for cache misses. The output cost, however, remains consistent at 12 yuan per million tokens. At present, the official App, web version, mini-program, as well as the DeepSeek API model, have all undergone simultaneous updates.