Imagine training a model with trillions of parameters without your hardware melting down. That is the challenge researchers tackled by optimizing the DeepSeek-V4 family on Ascend SuperPOD infrastructure. They did not just tweak settings; they built a complete system to handle the massive memory pressure and communication bottlenecks that usually cripple such large-scale efforts.

The core of their solution is a hierarchical framework that synchronizes model parallelism, computation, and low-level kernel execution. Most large language models rely on GPU clusters, but this team proved that Ascend NPUs can handle the load with serious efficiency. By orchestrating these complex layers, they achieved a 34.22% Model FLOPs Utilization rate. This represents a 2.93x improvement over open-source baselines, all while keeping the training process stable and reliable.

But efficiency is only half the story. The team used this optimized infrastructure to specialize the model for complex Operations Research tasks. They created a unique dataset combining domain resources with solver-verified synthetic documents, resulting in ten thousand high-quality samples. The outcome was a specialized model that outperformed GPT-5.4-Mini and the base DeepSeek-V4-Flash model in zero-shot reasoning. It achieved a Pass@1 score of 71.81%, proving that targeted post-training on alternative hardware can produce frontier-level performance for difficult mathematical modeling.

This work offers a clear blueprint for the future of large model training. It shows that with the right system-level optimizations, you can push trillion-parameter models to new heights on non-GPU hardware. More importantly, it demonstrates how domain-specific data can transform a general model into a specialist that beats the biggest players in the field.