SLAI T-Rex Optimizes DeepSeek-V4 Training with Ascend NPU

SLAI T-Rex is an end-to-end optimization framework for full-parameter post-training of trillion-parameter-scale MoE models on Ascend NPU SuperPOD.

Key achievements:

• Achieved 34.22% MFU — a 2.93x improvement over the open-source baseline
• Hierarchical optimization covering the entire stack: model parallelism, compute-communication orchestration, and low-level kernel execution

An Operations Research (OR) specialized model was also released. Built on DeepSeek-V4-Flash with a CPT·SFT pipeline, it was trained on 10,000 high-quality SFT samples including solver-verified synthetic data.

• Achieved zero-shot Pass@1 of 71.81%
• Outperformed GPT-5.4-Mini by +3.98pp and the base DeepSeek-V4-Flash by +11.27pp

The code is available on GitHub, and the model is available on ModelScope.