Huawei Ascend chips complete full-parameter post-training of DeepSeek-V4

A research team including Huawei Technologies announced it successfully used the firm's Ascend 910C chips to complete post-training of DeepSeek-V4-Pro, China's largest model to date at 1.6 trillion parameters, running on a cluster of at least 1,000 Huawei chips, according to a Shenzhen government social-media post relayed by SCMP.

The significance is in the distinction between inference and training. Chinese chipmakers have managed AI inference — running finished models to answer prompts — but have struggled with the far more complex compute- and communication-intensive work of training and refining a model's 'brain.' The team conducted 'full-parameter' post-training, meaning the entire architecture was updated without cutting corners. The government post likened it to adding 'complex flyovers and loops' to a previously 'one-way road,' instantly multiplying computational and communication demands.

Strategically, this is a notable leap for China's AI self-reliance under tightening US export controls, demonstrating that domestic silicon can handle post-training of a frontier-scale model. It dovetails with the week's other China-AI signals: Alibaba's open-weight Wan 2.7 video model winning the race 'OpenAI abandoned,' and a Hong Kong DeepSeek-V4-based model optimized for domestic chips.

Caveats: post-training is less compute-intensive than full pre-training, so this is a milestone rather than parity with NVIDIA-scale training; efficiency, yield, and cost versus Blackwell remain open questions. What to watch: whether Huawei can demonstrate from-scratch pre-training at scale, and how this shifts the US-China compute-decoupling trajectory.