Huawei's Ascend 910C chips achieve landmark post-training for 1.6-trillion-parameter DeepSeek model

A research team led by Huawei Technologies has successfully used the company's Ascend 910C chips to perform full-parameter post-training on DeepSeek's largest model to date, the DeepSeek-V4-Pro, which boasts 1.6 trillion parameters. The achievement, announced via a Shenzhen government social media post, involved a computing cluster of at least 1,000 Ascend 910C chips. Unlike basic inference—where a finished model answers prompts—post-training refines the model's entire architecture, teaching it to follow human instructions, safety rules, and specific tasks. The team described this as adding "complex flyovers and loops" to a previously one-way road, multiplying computational and communication demands several times.

This breakthrough marks a critical step for China's semiconductor industry, which has primarily focused on inference due to US sanctions restricting access to advanced training hardware. By successfully executing full-parameter post-training on such a massive model, Huawei demonstrates that Chinese chips can now handle the far more complex demands of model refinement. The project was a joint effort with the Shenzhen Loop Area Institute, the Shenzhen campus of Harbin Institute of Technology, and the Shenzhen Research Institute of Big Data. The Shenzhen government noted that this will "help enhance the self-reliance of China's AI industry chain," signaling a potential shift in the global AI hardware landscape.