Cambricon Announces Day 0 Support for DeepSeek V4.1 Flash

Chinese AI chip maker Cambricon announced on Sept. 10 that it has completed Day 0 adaptation for DeepSeek's latest open-source model V4.1 Flash based on the vLLM inference framework, allowing the model to run stably on Cambricon chips on the day of its release.

DeepSeek V4.1 Flash is the smallest model in DeepSeek's new architecture family, a 552B-parameter mixture-of-experts model. It features a new Causal-Encoder-Decoder structure, activating 8B parameters on the input side and 16B on the output side, with native multimodal visual understanding. DeepSeek says the model's cost is significantly lower than known models of similar size. The model also achieves a major breakthrough in KV cache compression, cutting HBM requirements to one-quarter and SSD requirements to one-eighth of the previous generation, with the KV cache reduced 437 times compared with the original model.

For the adaptation, Cambricon used its proprietary high-performance fused operator library Torch-MLU-Ops to accelerate model components such as Engram and Compressor, and used the BangC programming language to write optimized kernels for hot operators like sparse and compressed attention. At the framework level, Cambricon supports TP/PP/SP/DP/EP 5D hybrid parallelism, communication-computation overlap, low-precision quantization, and PD-disaggregated deployment in vLLM.

The move builds on Cambricon's NeuWare software ecosystem. On Sept. 8, the company joined the PyTorch Foundation as a top-tier Platinum member, gaining a seat on its governing board. So far, Cambricon has completed inference adaptation for five major Chinese-developed open-source models: GLM, DeepSeek, Qwen, Kimi and MiniMax.