DeepSeek新版本API价格下调寒武纪：对新模型DeepSeek

Core Insights - DeepSeek released the experimental version DeepSeek-V3.2-Exp on September 29, building on the previous version V3.1-Terminus with the introduction of DeepSeek Sparse Attention [1] - Cambrian announced on the same day that it has adapted to DeepSeek-V3.2-Exp and open-sourced the vLLM-MLU inference engine source code, emphasizing its commitment to the development of the large model software ecosystem [1] - Huawei Ascend quickly adapted and deployed DeepSeek-V3.2-Exp using the vLLM/SGLang inference frameworks, also open-sourcing all inference code and operator implementations for developers [1] Company Developments - Cambrian has been actively supporting mainstream open-source large models, including DeepSeek, as part of its long-term ecosystem development and technical accumulation [1] - Huawei's Ascend has achieved 0-day support for DeepSeek-V3.2-Exp, showcasing its rapid adaptation capabilities in the AI model landscape [1] Model Iteration - DeepSeek has been consistently advancing its model iterations, with the announcement of DeepSeek-V3.1 on August 21, followed by an update to DeepSeek-V3.1-Terminus on September 22 [1]