A 552B model, a quarter of the HBM, and a day-one Cambricon port

DeepSeek published the checkpoint for DeepSeek-V4.1-Flash, and Cambricon says the model ran on its accelerators the same day. The integration target was vLLM, the open-source inference stack operators already standardize on, rather than a private fork of it.

News like this arrives dressed as model news. The claim worth testing is narrower: a chip vendor shipped a working serving path on launch day and described the pieces in enough detail to be judged on engineering rather than on announcement language.

What Cambricon actually shipped for DeepSeek-V4.1-Flash

The deliverable is a fused operator library, a set of hand-written kernels and a serving configuration. Cambricon engineers used Torch-MLU-Ops to accelerate structures the company names Engram and Compressor, and wrote optimized BangC kernels for sparse and compressed attention. Inside vLLM, Cambricon says its path supports five-dimensional mixed parallelism across tensor, pipeline, sequence, data and expert dimensions, with communication overlapped against compute, plus low-precision quantization and prefill-decode disaggregation. The whole stack sits on NeuWare, Cambricon's software platform of long standing.

That covers the parts of a serving stack that decide whether a cluster stays busy. Demand for that capacity keeps rising, with Nvidia's market cap back above $3 trillion on infrastructure spending. The list does not include a number. The announcement carries no throughput figure, no latency figure and no comparison against another accelerator, so same-day availability is a scheduling claim rather than a performance one.

Cambricon joined the PyTorch Foundation as a Platinum member days earlier, which puts it on the governing board alongside other hardware and cloud contributors. Read together with the Day-0 port, the sequence points to upstream investment rather than a one-off patch.

The memory math behind the same-day port

DeepSeek-V4.1-Flash is the smallest member of DeepSeek's newer architecture family. It is a 552-billion-parameter mixture-of-experts model that activates about 8 billion parameters on the input side and 16 billion on the output side, and it handles vision natively. The layout is a Causal-Encoder-Decoder. A large parameter count over a small active slice is now routine among open MoE launches, as with Xiaohongshu's 280B dots3-note, which activates 16B.

The compression is what accelerator vendors care about. DeepSeek says KV-cache compression brings HBM demand to roughly one-quarter and SSD demand to roughly one-eighth of the previous generation, with cache size described as hundreds of times smaller than the first-generation design. Those cuts land against a memory shortage already pushing hardware prices upward.

Metric	DeepSeek-V4.1-Flash vs prior generation
HBM demand	roughly one-quarter
SSD demand	roughly one-eighth
KV cache size	described as hundreds of times smaller than the first-generation design

On a dense serving node, the KV cache caps how many concurrent sequences fit in memory, and those nodes already strain HBM budgets. Saving three-quarters of it is either more users per card or fewer cards per deployment, which is why a smaller cache shows up first as enthusiasm among chip vendors. These are DeepSeek's figures, stated for the model rather than measured on Cambricon hardware, and no independent check has been published. More memory capacity translates directly into more concurrency, as Hugging Face's MI455X tests measured at roughly three times the request throughput.

Five model lines, one recipe

Cambricon lists Day-0 or production inference support across five Chinese model lines: GLM from Zhipu AI, DeepSeek, Qwen from Alibaba, Kimi from Moonshot AI, and MiniMax.

The wording announces less than it looks. "Day-0 or production" covers launch-day readiness and support bolted onto models that shipped months earlier, and the list does not separate the two.

Model line	Developer	Cambricon's stated status
GLM	Zhipu AI	Day-0 or production inference support
DeepSeek	DeepSeek	Day-0 or production inference support
Qwen	Alibaba	Day-0 or production inference support
Kimi	Moonshot AI	Day-0 or production inference support
MiniMax	MiniMax	Day-0 or production inference support

For cluster operators, that distinction is the whole value of the claim. A launch-day port means a documented recipe exists when the checkpoint drops. A retrofit means waiting for the vendor to catch up after the fact.

Closed forks versus public frameworks

Caixin Global noted the same-day enablement as a marker that domestic model-to-chip software loops are shortening, from multi-month private ports toward launch-day readiness on public frameworks operators already run. The short being measured is a software short, not a silicon one. A vendor that can follow upstream vLLM releases carries a cheaper support burden than one maintaining a closed fork of every model it serves.

What the announcement leaves open is whether the pattern scales. DeepSeek-V4.1-Flash is the smallest member of its family, and Cambricon's own framing points to the next test: larger DeepSeek variants and multimodal serving mixes. Porting the small model in a day is a good sign. It is not yet proof that the same team absorbs a bigger checkpoint, or a workload serving vision and text together, on the same schedule.

Day-0 has become a claim several vendors can make. The one that keeps making it as the checkpoints get larger is the one worth watching.