Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Moore Threads completes Kimi K3.1 Pro training adaptation, holding 400 stable steps on 16 GPUs

Moore Threads has completed a training adaptation for Kimi K3.1 Pro and held 400 consecutive training steps stable on two machines with 16 GPUs, according to a report carried by East Money. The run moves domestic accelerator validation from inference into the harder problem of multi-node training.

Published

Moore Threads has finished adapting its hardware and software stack for training Kimi K3.1 Pro, according to a report carried by East Money. The run used two machines and 16 GPUs and held stable for 400 consecutive training steps.

The substance sits in two phrases: training adaptation and continuous stability. The first means the operators, communication libraries and framework hooks a large model depends on now work on Moore Threads' stack. The second means the job did not collapse mid-run from communication faults, memory overflow or numerical issues.

Training is materially harder to port than inference. Inference requests can be split and served independently, while training requires every device to stay in lockstep at each gradient synchronization point, so a single slow or failing card drags the whole cluster. Sixteen GPUs across two nodes is a modest footprint, but it is enough to expose cross-node communication, collective-communication efficiency and long-run stability.

The 400-step figure should be read precisely. It demonstrates stability over a continuous stretch of training, not convergence of a finished model. Production runs stretch across tens of thousands of steps and hundreds or thousands of accelerators, so this adaptation is closer to proof that the stack can run than a position on a performance leaderboard.

The competitive rhythm among domestic accelerator vendors is shifting from “can it infer” to “can it train”. Over the past two years, chip makers and model teams have announced a steady stream of adaptation work, which serves double duty: model builders get a second source of compute, and chip vendors get real workloads to tune against.

What would matter next is more specific numbers — scaling efficiency on larger clusters, end-to-end training throughput, framework compatibility and recovery behaviour over long runs. Those metrics say far more about how close a domestic training stack is to production than the phrase “adaptation completed” ever will.

Why it matters

Training adaptation, not inference, is the harder test for domestic accelerators, and a stable multi-node run is the first real signal that the software stack can carry large-model workloads.

Moore ThreadsKimiGPUTraining
Back to AI Daily

Nearby Updates

All