Cambricon announced the successful completion of adapting the DeepSeek-V4.1-Flash model as part of the Day-0 program based on the open inference stack vLLM. This means that the recently released model can run stably on Cambricon accelerators on the same day DeepSeek published the checkpoint.
This announcement highlights the synergy between chip and software: simultaneous availability of the model and the chip, as well as the maturity of Cambricon's approach combining NeuWare and vLLM. This is not a repetition of information about the general availability of the model itself.
DeepSeek-V4.1-Flash represents the smallest variant of the new DeepSeek architectural line. It is a mixture-of-experts model with 552 billion parameters, which activates about 8 billion parameters on the input side and 16 billion on the output side, and also possesses built-in multimodal vision understanding. DeepSeek notes the use of the Causal-Encoder-Decoder structure and aggressive KV cache compression, which reduces HBM requirements by approximately one quarter and SSD requirements by one eighth compared to the previous generation, while the cache size is hundreds of times smaller than the first version.
These memory savings are critically important for accelerator manufacturers who aim to control inference latency, DRAM volume, and storage levels on dense servers that are already straining HBM budgets.
For its part, Cambricon engineers used the Torch-MLU-Ops unified operator library to accelerate structures such as Engram and Compressor, and wrote optimized kernels in BangC for sparse and compressed attention hotspots. Within vLLM, Cambricon reports support for five-dimensional mixed parallelism, covering tensor, pipeline, sequential, batch, and expert parallelism, with communication and computation overlap, low-precision quantization, and prefix and decoding splitting to boost end-to-end throughput. This stack operates on NeuWare, the company's long-term software platform.
Several days ago, Cambricon joined the PyTorch Foundation as a platinum member, taking a seat on the governing board alongside other major hardware and cloud service developers, signaling deeper investments in the open learning and inference ecosystem.
Cambricon now supports Day-0 inference for five major Chinese model lines: GLM, DeepSeek, Qwen, Kimi, and MiniMax, representing Zhipu AI, DeepSeek, Alibaba, Moonshot AI, and MiniMax, respectively. Caixin Global specifically noted the ability to include Cambricon on the same day as an indicator of shortening internal development cycles for models and chips in China—a shift from multi-month closed ports to readiness for launch on public frameworks that are already standardized by operators.
The practical takeaway for cluster operators is co-availability: when a cutting-edge open model is released, Cambricon clients can use the documented vLLM recipe instead of waiting for a closed fork. The next test for the software ecosystem will be whether this Day-0 scheme applies to larger DeepSeek variants and multimodal serving mixtures.
