Cambricon Announces Full Series Model Day0 Compatibility and Open-Sourcing of Optimized Code for DeepSeek-V4

Cambricon announced the completion of Day0 compatibility for the entire series of DeepSeek-V4 models. Based on the vLLM inference framework, it covers the 285B parameter Flash version and the 1.6T parameter Pro version.

By optimizing sparse attention and compressed structures with its self-developed Torch-MLU-Ops operator library, the model can run stably on Cambricon hardware on the day of release. The related code has been open-sourced to GitHub.