Within a Day, China's Largest Open Model Was Running on Chinese Accelerators

Huawei announced at 08:17 UTC on 28 July that its Ascend stack had achieved what it calls "0day" adaptation for Kimi K3 — roughly nineteen hours after Moonshot uploaded the weights. The claim is that China's largest open model runs on Chinese accelerators with no Nvidia hardware anywhere in the stated path.

The stated stack

Training via MindSpeed MM on Atlas 800 A3 and Atlas 900 A3 SuperPoD systems, with operator-level, parallel-acceleration and memory optimisations. Inference through vLLM Ascend and SGLang. Huawei states that its Ascend 950 super nodes support the full numerical range required — FP8, MXFP8 and MXFP4. That last format is the one that matters: MXFP4 is exactly the weight format Kimi K3 ships in, per its model card.

What '0day' does and does not mean

It is Huawei's own marketing term, and it describes adaptation availability — that the software path exists and runs. It is not a performance claim. No benchmark, no token throughput, no latency and no cost-per-token figure was published. Ascend 950's support for MXFP4 is asserted rather than demonstrated. Nothing here establishes that running a 2.8-trillion-parameter model on this hardware is economically competitive, only that it is possible.

Why the pairing is the point

Take the week's other stories together. A Chinese lab released frontier-scale open weights; a Chinese manufacturer began building the lithography tools that make domestic accelerators; and a Chinese vendor demonstrated the model running on its own silicon inside a day. Each link is separately unremarkable. The chain is the argument China has been making for three years, assembled in public over roughly thirty-six hours — and it is the same argument that took 14.65% off SK Hynix the same morning.

Sourcing caveat

The announcement went out through Huawei's WeChat official account, which is not openly fetchable, and was carried by a single Chinese technology outlet at the time of writing. The hardware names check out against Huawei's own developer site. Treat the capability claim as self-reported and expect independent confirmation to arrive later than the headline.