On April 24, Cambricon made an announcement that it has successfully achieved 'Day 0' compatibility with DeepSeek-V4. DeepSeek-V4 is the latest open-source large-scale model developed by DeepSeek Company. This compatibility is built upon Cambricon's proprietary NeuWare software ecosystem and the vLLM inference framework. In a move that underscores its commitment to open collaboration, Cambricon has also open-sourced the adaptation code to the GitHub community at the same time.

Leveraging its in-house high-performance fused operator library, Torch-MLU-Ops, Cambricon has accomplished specialized acceleration for key modules such as Compressor and mHC. Furthermore, the company has employed the BangC programming language to craft highly optimized kernels, thereby fully harnessing the potential of the underlying hardware.

In tandem with these developments, the preview version of DeepSeek-V4 has been rolled out and open-sourced. This version boasts an impressive ultra-long context capacity of up to one million words, along with exceptional Agent capabilities and inference performance. It's worth noting that Hygon DCU has also concurrently completed its 'Day 0' adaptation for DeepSeek V4, marking another significant milestone in the field.