SenseTime has released the final checkpoint of SenseNova-Loop-9B, with Cambricon reporting day-0 support on its MLU accelerators.
Day-0 support has become a meaningful signal in the Chinese market. It means a model can run on domestic silicon at launch rather than months later, which is the practical test of whether a chip vendor's software stack is keeping pace with model architecture changes.
At 9 billion parameters the model is deliberately mid-sized: large enough for reasoning-heavy tasks, small enough to serve on a single accelerator with quantisation - the segment where domestic inference demand is concentrated.
Both companies attached caveats around memory requirements and reported benchmarks, which is appropriate. Vendor-published numbers on unpublished evaluation protocols should be treated as directional.
The broader pattern is coordination: model releases in China increasingly arrive with a chip partner already integrated, compressing the gap between a new architecture and deployable capacity.
