SLAI T-Rex hits 34.22% MFU post-training DeepSeek-V4 on Ascend

A new paper posted to Hugging Face's daily papers feed is worth reading as a systems result more than a model result. A team calling itself SLAI reports it did full-parameter post-training of a trillion-parameter MoE, the DeepSeek-V4 family, entirely on Huawei's Ascend NPU SuperPOD, and hit 34.22% Model FLOPs Utilization. Their own baseline number for the open-source recipe on the same hardware was roughly a third of that, so they are claiming a 2.93x speedup from their optimization stack.

The interesting bit is what the stack actually does. The authors describe a three-level framework: model-level parallelism, computation-communication orchestration to hide the non-overlapped comms overhead that tends to kill MFU on distributed MoE training, and low-level kernel work tuned to the Ascend NPU. None of those ideas are new on their own, but bundling them into a public recipe that gets a trillion-parameter post-training run into the mid-thirties MFU range without any Nvidia silicon in the loop is the news.

On top of the systems work they also fine-tuned a domain-specialized DeepSeek-V4-Flash variant on 10,000 samples covering four task categories and three problem representations, aimed at Operations Research. It scores 71.81% zero-shot Pass@1 on their OR evaluation, which the paper puts at 3.98 points above GPT-5.4-Mini and 11.27 points above the base DeepSeek-V4-Flash. The training data is described as solver-verified synthetic optimization documents, which is a sensible way to make an OR fine-tune actually correct rather than plausible-sounding.

Take the specifics as reported, not settled. The MFU figure is self-published, the OR benchmark is a narrow slice rather than general reasoning, and the write-up as summarized does not pin down the SuperPOD scale, the NPU-hours consumed, or the wall-clock time behind the 2.93x claim. Independent reproduction outside the Ascend stack is unlikely for obvious reasons.

The forward-looking read is that the fully-Chinese frontier post-training stack is starting to publish credible recipes rather than only marketing numbers, and Ascend appears to be moving from an inference story toward a training one. If you are watching hardware-diversification risk in your model roadmap, or you sell into operations research, this is a data point worth filing.