Huawei-led team claims it post-trained DeepSeek's 1.6-trillion-parameter model — 1,000 Ascend 910C chips used in training

A Huawei-led research group has claimed successful full-parameter post-training of DeepSeek's V4-Pro, a 1.6-trillion-parameter model, using a cluster of at least 1,000 Huawei Ascend 910C chips. This d... Read original →

Huawei’s claim of full-parameter post-training on a 1.6T model using 1,000 Ascend 910C chips matters less for raw performance and more for closing China’s domestic AI training stack gap. Technically, stable large-scale training would force rapid CANN software upgrades and shift Ascend from inference-only to genuine training relevance. Geopolitically, this counters U.S. export controls by reducing reliance on CUDA, though pre-training from scratch remains out of reach. NVIDIA will likely tighten software locks on China-specific chips like H20 and accelerate Grace-Hopper adoption elsewhere. Over the next 18 months, China’s AI infrastructure will bifurcate—training on homegrown hardware, inference on heterogeneous accelerators—but without transparent benchmarks, global developers won’t trust Ascend, risking an isolated, inward-looking ecosystem.