A research team led by Huawei Technologies has claimed to have completed full-parameter post-training of DeepSeek’s V4-Pro model, a 1.6 trillion-parameter system, on a cluster of at least 1,000 Ascend 910C chips. The announcement, made via a Shenzhen municipal government social media post on June 6, represents a meaningful step forward in China’s effort to demonstrate that its domestic AI hardware can handle not just inference but the far more demanding task of training-class workloads.
The research team includes Huawei Technologies, the Shenzhen Loop Area Institute, the Shenzhen campus of Harbin Institute of Technology, and the Shenzhen Research Institute of Big Data.
From Inference to Training: Why This Milestone Matters
Previous coverage of China’s domestic AI hardware has focused primarily on inference performance, the ability to run a finished model and generate responses. On that front, Huawei’s Ascend chips have demonstrated roughly 60% of Nvidia H100 inference performance in earlier DeepSeek testing, a gap that Chinese firms have partially closed through software optimization.
Post-training is a fundamentally different and far more demanding challenge. It involves updating every weight in the model — in this case, across 1.6 trillion parameters — through techniques such as supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). The computational and communication demands are exponentially higher than inference, requiring all-reduce operations that synchronize gradient updates across thousands of chips simultaneously, a workload that has historically exposed the weaknesses in Huawei’s CANN software stack.
This distinction matters enormously for China’s AI self-reliance ambitions. A chip ecosystem that can only run finished models remains dependent on foreign hardware for the training pipelines that produce those models in the first place.
A Reversal from August 2025
The significance of this claim becomes clearer in the context of where things stood less than a year ago. In August 2025, DeepSeek reportedly failed to complete a single successful training run on its R2 model using Ascend chips, citing unstable performance, slow chip-to-chip interconnects, and gaps in Huawei’s CANN software stack.
DeepSeek-V4-Pro, released in April 2026, was the first DeepSeek model built around Ascend hardware from the outset, with its pre-training corpus exceeding 32 trillion tokens. The new post-training claim suggests that the software and hardware co-optimization work that began with V4-Pro has now extended to the full training pipeline.
Important Caveats
The claim comes with significant limitations that warrant caution. Tom’s Hardware flagged several conspicuous gaps: the announcement is silent on training duration, cluster utilization rates, and any performance comparison with equivalent Nvidia hardware. No independent benchmarks were released alongside the claim, and DeepSeek has not issued any statement confirming or elaborating on the results.
The absence of independent verification means this should be treated as a claimed milestone rather than a confirmed one. The Shenzhen government’s framing, that the work “will help enhance the self-reliance of China’s AI industry chain,” suggests the announcement carries political as well as technical significance.
The Software Stack Behind the Claim
The post-training run was made possible not just by the 1,000-chip Ascend 910C cluster but by Huawei’s CANN (Compute Architecture for Neural Networks) software stack, which has been the primary bottleneck in previous attempts to run training-class workloads on Ascend hardware. The August 2025 failure with DeepSeek R2 was attributed in part to gaps in CANN’s support for the communication primitives required during distributed training, specifically, the all-reduce operations that synchronize gradient updates across thousands of chips simultaneously.
The South China Morning Post’s reporting on the new claim suggests that the Huawei-led team has made substantial progress on these software-level challenges, though the details of what specifically changed in the CANN stack between August 2025 and June 2026 have not been publicly disclosed. The involvement of the Shenzhen Research Institute of Big Data and the Shenzhen campus of Harbin Institute of Technology alongside Huawei suggests a collaborative engineering effort rather than a single-company breakthrough.
Implications for the Chip War
Even with those caveats, the direction of travel is clear. Chinese tech giants have been scrambling to secure Huawei Ascend chips since DeepSeek V4-Pro triggered a supply crunch, and Huawei’s Ascend 950 production is expected to scale to 750,000 units in 2026. If the post-training claim holds up under scrutiny, it would mark a critical inflection point: China moving from running AI models on domestic chips to building and refining them entirely on domestic silicon, closing the most important remaining gap in its AI self-reliance stack.