DeepSeek V4 Pro post-trained on a thousand Huawei chips without any NVIDIA hardware
A joint research team led by Huawei says it completed full-parameter post-training of DeepSeek’s 1.6-trillion-parameter V4-Pro model using 1,000 Huawei Ascend 910C accelerators. No NVIDIA silicon was involved. The Shenzhen municipal government announced the result, calling it a milestone in China’s push toward AI self-sufficiency under tightening US export controls.
The team included Huawei, the Shenzhen Loop Area Research Institute, Harbin Institute of Technology’s Shenzhen campus, and the Shenzhen Research Institute of Big Data. The work was carried out on a computing cluster of at least 1,000 Ascend 910C processors, according to the South China Morning Post.
Full-parameter post-training is not fine-tuning. Every weight in the model was updated, rather than bolting on a thin adapter layer. Think of pre-training as teaching a model to speak by absorbing massive text corpora. DeepSeek V4-Pro’s pre-training corpus exceeds 32 trillion tokens. Post-training then teaches the model to follow instructions, adhere to safety rules, and handle specific tasks. Completing this stage on domestic Chinese silicon is a genuine step change.
Why the Ascend 910C matters
The Ascend 910C is Huawei’s flagship AI accelerator, built on SMIC’s second-generation 7nm process using Huawei’s proprietary Da Vinci architecture. Huawei claims it delivers performance comparable to NVIDIA’s H100. Earlier DeepSeek testing put the 910C at roughly 60% of an H100’s inference performance. That is competitive for running models, but training is where Chinese silicon has consistently fallen short.
Chinese chips have been decent at inference for a while. You load a finished model, ask questions, get answers. Training is the hard part. Recalculating billions of weights across enormous datasets demands sustained compute, high-bandwidth interconnect, and mature software tooling. NVIDIA’s CUDA ecosystem has been the moat. This result suggests that moat is narrowing.
The Shenzhen government framed it as going from a one-way road (input question, output answer) to adding complex overpasses and loops. The compute and communication demands increased several-fold compared to inference-only workloads.
What this does not prove
Post-training is the lighter stage. Pre-training a frontier model from scratch, which is where the bulk of compute cost lives, remains unproven on Ascend hardware. Tom’s Hardware correctly noted that this result does not demonstrate the chips can handle that heavier job. The gap between “we tuned an existing frontier model” and “we built one from nothing on domestic silicon” remains significant.
That said, post-training is where much of the practical model quality comes from. A model that pre-trains on NVIDIA but post-trains on Ascend is still a model that partially escapes NVIDIA’s supply chain. For Chinese AI labs under export restrictions, that partial independence matters.
The geopolitical context
US export controls have progressively tightened around NVIDIA’s most advanced chips. China banned foreign AI chips from state-run data centers earlier this year. NVIDIA CEO Jensen Huang visited Beijing to discuss workarounds. The Chinese government ordered major tech companies to stop buying NVIDIA hardware outright.
DeepSeek itself has been at the center of this tension. The company launched V4-Pro, its 1.6-trillion-parameter flagship, on Huawei chips as US officials escalated accusations of AI-related intellectual property theft. The fact that DeepSeek is now demonstrating post-training results on Ascend hardware is not accidental. It is the intended outcome of years of industrial policy and billions of yuan in subsidies.
Hacker News readers debated the result extensively on Hacker News, where the story hit 328 points and 165 comments. The discussion split between those who see a genuine narrowing of the hardware gap and those pointing out that post-training is still far from full pre-training independence.
What to watch next
The next signal to watch is whether any Chinese lab claims full pre-training of a frontier-scale model on Ascend hardware. That would be the real inflection point. Until then, this result is best understood as proof that Huawei’s silicon can handle the second-heaviest stage of the AI pipeline at scale. That is more than most people expected even a year ago.