Chinese AI company DeepSeek released the latest model DeepSeek-V3.2-Exp, which provides optimization for China's Ascend chips and its CANN (Compute Architecture for Neural Networks) computing architecture. This represents the shift of DeepSeek's center of gravity, ensuring that advanced models can be operated in China's AI accelerators, and no longer blindly rely on the NVIDIA CUDA ecosystem.
With DeepSeek-V3.2-Exp, China quickly integrates new models for the Sheng Teng team and the related vLLM-Ascend community. In vLLM-Ascend storage, new projects outline custom installation steps and core packaging for Sheng Teng NPUs to support new models. The CANN team also issued inference deployment guidelines to enable new models to be deployed on the chips.
Other Chinese chip suppliers support, including Cambrian updates to the vLLM-MLU branch to be compatible with DeepSeek-V3.2-Exp, claiming that its inference engine is combined with the new model's sparse attention mechanism to reduce long-sequence processing costs. Haiguang also said that its DCU accelerator is adjusted through the DTK software stack to achieve "zero-wait" deployment.
At the same time, the reasoning framework SGLang confirms that DeepSeek-V3.2-Exp can support multiple backends (including Shengren), and DeepSeek's GitHub description implies that the new model can achieve comparable compatibility with vLLM when it is launched. DeepSeek also mentioned the high-level language TileLang and CUDA cores, and suggested that researchers should focus on TileLang when developing prototypes. In fact, this means that the same model file can be deployed between an NVIDIA GPU and a Chinese accelerator with a small amount of adjustments.
The rapid adoption of this is highlighting that China's AI ecosystem is preparing for the future of naturally obtaining NNVIDIA hardware. Although NVIDIA's CUDA maintains its leading position in the field of training and reasoning, the latest version of DeepSeek is a small number of Chinese companies that provide optimized products for non-CUDA software stacks on the first day of its release.
Extended reading: DeepSeek-V3.2-Exp Published! Improve the efficiency of reasoning, and the price of API is cut at least half