Moonshot AI released the open weights of its Kimi K3.1 Pro model on 4 November. Huawei's Ascend team now says LoRA fine-tuning of it works on a 128-card Atlas 950 SuperPoD. LoRA trains a small add-on layer rather than every weight. Ascend says training loss fell from about 1.85 to about 0.27 over more than 500 steps.
The model has about 2.8 trillion parameters, and its weights alone take about 5.6TB. Training needs much more memory for gradients, optimizer state and long inputs. Ascend says it shards these across cards and spreads the 896 experts over them, so each card holds less. It says Swap Core saves the results of costly matrix steps, so the backward pass skips repeating them. Ascend credits this with a 15% average speed gain. Pandaily reported no absolute speeds and no outside check of the 15%.
Ascend presents this as a working route; the evidence so far is its own loss curves. They show training runs, not how good the tuned model is. Running at this size is a memory problem first, so "it runs" leaves the numbers buyers need unpublished. Buyers should ask for speed, run time and cost. They should also ask for comparisons with other chips.