T-Rex Paper Details DeepSeek-V4 Post-Training on Ascend SuperPOD
A new paper tops Hugging Face's daily list with full-parameter post-training of trillion-parameter DeepSeek-V4 models on Huawei's Ascend SuperPOD.
A 65-author paper titled "SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD," posted to arXiv on July 22, climbed to the top of Hugging Face's daily papers ranking on July 23. The work documents something the AI-infrastructure world has watched for closely: full-parameter post-training of frontier-scale mixture-of-experts models — the DeepSeek-V4 family, whose largest member weighs in around 1.6 trillion parameters — carried out entirely on Huawei's Ascend NPU SuperPOD rather than Nvidia GPUs. Coverage by Tom's Hardware puts the cluster at roughly 1,000 Ascend 910C chips.
Engineering around the hard parts
Full-parameter post-training at trillion-parameter scale stresses exactly the areas where non-Nvidia stacks have historically struggled: memory pressure, communication that cannot be hidden behind computation, and inefficient kernels. The team's answer is a hierarchical optimization framework spanning model parallelism, computation-communication orchestration, and low-level kernel execution. The reported outcome is 34.22 percent Model FLOPs Utilization with stable training — a 2.93x improvement over the open-source baseline recipe for the same hardware.
The paper pairs the systems work with a domain case study. Using continued pre-training and a 10,000-sample supervised fine-tuning set spanning four operations-research task types, the team specialized DeepSeek-V4-Flash into a solver-grounded reasoning model that reaches 71.81 percent zero-shot Pass@1 on their evaluation — 11.27 points above the base model and, by their measurement, 3.98 points above GPT-5.4-Mini.
Why it matters
Inference on Ascend chips is by now routine in China; frontier-scale training is the harder claim, and this paper is among the most detailed public accounts of it. If trillion-parameter MoE models can be post-trained at competitive efficiency on domestic silicon, US export controls lose some of their bite at precisely the stage — turning strong base models into specialized, deployable ones — where most applied AI work now happens. The community reaction, pushing the paper past everything else on Hugging Face's daily list, suggests researchers read it the same way: as evidence that the Nvidia-free training stack is maturing faster than skeptics assumed.