Huawei Ascend 910C chips complete DeepSeek-V4-Pro post-training
Synopsis
For the first time, Huawei's Ascend 910C chips have completed full-parameter post-training on DeepSeek's largest-ever 1.6-trillion-parameter model using a 1,000-chip cluster — a concrete sign that China's domestic AI hardware is moving beyond inference into frontier model training.
Key Takeaways
Huawei's Ascend 910C chips completed full-parameter post-training for DeepSeek-V4-Pro , announced on 5 June 2026 .
The model has 1.6 trillion parameters , making it DeepSeek's largest to date.
Training ran on a cluster of at least 1,000 Huawei Ascend 910C chips.
The project was jointly conducted by Huawei , Shenzhen Loop Area Institute , Harbin Institute of Technology (Shenzhen campus) , and Shenzhen Research Institute of Big Data .
The achievement marks a shift from inference-only domestic compute to full model post-training, directly relevant to US chip export control strategy.
The Shenzhen government stated the project 'will help enhance the self-reliance of China's AI industry chain.'
A research consortium that includes Huawei Technologies has successfully used the company's Ascend 910C chips to complete post-training for the DeepSeek-V4-Pro large language model — a milestone that signals China's domestic semiconductor ecosystem is pushing beyond basic AI inference into the far more demanding territory of model training, amid escalating US export controls on advanced chips.
What happened
According to a Friday, 5 June 2026 social media post from the Shenzhen municipal government, the team ran DeepSeek's largest model to date — a behemoth with 1.6 trillion parameters — on a computing cluster powered by at least 1,000 Huawei Ascend 910C chips. The team conducted 'full-parameter' post-training, meaning the model's entire architecture was updated and refined without shortcuts or approximations.
The collaborative effort was jointly led by Huawei, the Shenzhen Loop Area Institute, the Shenzhen campus of Harbin Institute of Technology, and the Shenzhen Research Institute of Big Data.
Why it matters
The distinction between inference and training is critical. Inference — running a finished model to respond to prompts — is computationally modest. Training, and particularly post-training, is exponentially harder: it demands chips to handle massive parallel communication loads while simultaneously updating billions of parameters. Chinese chipmakers have historically struggled at this stage, with domestic hardware largely confined to inference workloads.
The Shenzhen government post described the leap vividly: where inference is 'much like building a one-way road for the model — input a question, output an answer,' post-training 'added complex flyovers and loops to that one-way road, instantly multiplying the computational and communication demands by several times.' The project enabled the model to self-reflect and adjust, rather than simply respond.
The competitive backdrop
The achievement arrives as Washington continues to tighten restrictions on the export of high-end Nvidia GPUs and related hardware to China. That pressure has accelerated state-backed investment in domestic alternatives, with Huawei's Ascend series emerging as the primary challenger to Nvidia's dominance in AI compute. Until now, however, Ascend chips had not been publicly demonstrated completing full-parameter post-training on a frontier-scale model of this size.
The DeepSeek-V4-Pro model's 1.6 trillion parameters make it one of the largest language models publicly associated with domestic Chinese compute infrastructure. Full-parameter post-training at this scale — without relying on foreign silicon — is a concrete data point in China's broader push for AI supply-chain self-reliance.
What's next
The research post stated the project 'will help enhance the self-reliance of China's AI industry chain.' The immediate question for the industry is whether the cluster's performance translates into reproducible, cost-competitive training runs that can challenge Nvidia-powered setups at commercial scale. Analysts and hardware engineers will be watching for peer-reviewed benchmarks and whether the collaboration leads to wider deployment of Ascend 910C clusters across China's AI labs and cloud providers.
Point of View
Where the compute bar is lower. Completing full-parameter post-training on a 1.6-trillion-parameter model on domestic silicon directly challenges that calculus. What mainstream coverage tends to underplay is the communication fabric problem: post-training at this scale is as much a networking challenge as a raw compute one, and Huawei clearing that bar suggests the Ascend ecosystem's interconnect stack is maturing faster than many Western analysts assumed. The key unknown remains cost and reproducibility — a single successful run backed by government institutions is not the same as a commercially viable, repeatable pipeline. The companies most exposed to watch are Nvidia, whose H-series lock-in in China is eroding, and Chinese cloud providers who must now decide how aggressively to migrate training workloads to domestic hardware.
NationPress
22 Jul 2026
Frequently Asked Questions
What did Huawei's Ascend 910C chips achieve with DeepSeek?
Huawei's Ascend 910C chips were used to complete full-parameter post-training for the DeepSeek-V4-Pro model, which has 1.6 trillion parameters. The training ran on a cluster of at least 1,000 Ascend 910C chips, according to the Shenzhen municipal government.
What is the difference between AI inference and post-training?
Inference is the process of running a finished AI model to answer user prompts — computationally modest and already supported by Chinese chips. Post-training is the far more demanding process of refining a model's behaviour using human instructions and safety rules, requiring exponentially more compute and chip-to-chip communication.
Why is this significant for China's semiconductor industry?
Chinese chipmakers had previously been limited mainly to inference workloads. Successfully completing full-parameter post-training on a frontier-scale model using only domestic silicon marks a concrete step toward AI supply-chain self-reliance, especially as US export controls restrict access to Nvidia's advanced GPUs.
Who conducted the DeepSeek post-training research with Huawei?
The project was a joint effort involving Huawei, the Shenzhen Loop Area Institute, the Shenzhen campus of Harbin Institute of Technology, and the Shenzhen Research Institute of Big Data, as announced on 5 June 2026.
How does this affect Nvidia's position in the Chinese AI market?
Nvidia has dominated AI training hardware globally, including in China before export restrictions tightened. Huawei demonstrating post-training capability at the 1.6-trillion-parameter scale on domestic chips increases competitive pressure on Nvidia and may accelerate the migration of Chinese AI labs toward Ascend-based infrastructure.
