China-based AI company DeepSeek has launched its DeepSeek-V3.2-Exp model, a new step toward reducing dependency on the NVIDIA CUDA ecosystem. The model has been optimized for Huawei’s Ascend accelerators and the company’s CANN software stack since day one. This accelerates the process of China-based technology companies becoming independent of US-based hardware and software ecosystems.

DeepSeek announced the model on September 29, 2025, and shared the code and checkpoints on Hugging Face along with a technical report. The company describes V3.2-Exp as “an intermediate step in the transition to our next-generation architecture, designed to reduce costs in long-context inference.” The model stands out for its sparse attention mechanism, which reduces memory and processing power requirements while maintaining output quality.

Huawei’s Ascend team and the broader vLLM-Ascend community quickly adopted the model. Specific operator installation steps and core packaging instructions have been published for its execution on Ascend NPUs. The CANN team also shared with users the inference recipe, which can be run directly on Huawei hardware.

DeepSeek is supported by Chinese manufacturers

Not only Huawei, but other Chinese chip manufacturers have also quickly joined the process. Cambricon updated its version of vLLM-MLU for its own accelerators, announcing that the model’s sparse attention structure provides cost advantages for long sequences. Hygon announced that it has made its DCU accelerators ready for deployment with “zero latency” via the DTK software stack.

It has also confirmed V3.2-Exp support on various hardware, including SGLang and Ascend. DeepSeek’s notes on GitHub indicate that the model offers performance equivalent to vLLM. The company mentions both TileLang and CUDA kernels, inviting researchers to use TileLang for prototyping. This approach allows the same model to be run on both NVIDIA GPUs and Chinese accelerators with only minor modifications.

These developments clearly demonstrate that China’s AI ecosystem is preparing for a future where access to NVIDIA hardware is no longer guaranteed. While CUDA remains dominant globally, DeepSeek’s new model is one of the few large-scale Chinese initiatives to be optimized for non-CUDA platforms from day one.

The coordinated efforts of Huawei, Cambricon, and Hygon demonstrate that companies are taking Beijing’s “technological dominance” goals seriously. Chinese manufacturers are no longer content with simply adapting to NVIDIA-based systems; they are positioning their own hardware and software platforms as first-class targets.