Huawei Says 128 Cards Tuned a 2.8-Trillion-Parameter Kimi Model

Huawei says it has adapted a 2.8-trillion-parameter AI model on a 128-card accelerator system, according to Pandaily, a test of its hardware's ability to handle very large models. ^[1]

The reported run used Moonshot AI's Kimi K3.1 Pro on an Atlas 950 SuperPoD, with a technique called LoRA that adjusts a model for a particular task without retraining all its original weights. Pandaily also attributes an average 15 percent performance gain to Huawei's Swap Core caching. That figure is Huawei's own and has not been independently verified, and the measurement, baseline and workload have not been established. ^[1]

A successful run would give developers evidence that Huawei's systems can handle one useful form of adaptation at this scale. It would not mean Huawei trained Kimi from scratch, or establish how quickly the system answers ordinary users' requests.

What the giant number means

Kimi K3.1 Pro's published specifications help explain the task. Moonshot AI's model card lists 2.8 trillion total parameters, the numerical values that encode what a model has learned. But it says 104 billion parameters are activated for each token, a unit of text the model processes. ^[4]

That is about 3.7 percent of the total parameter count. The model uses a mixture-of-experts architecture, which selects portions of the model to do the work for each token rather than activating the entire model at once. ^[4]

LoRA, short for low-rank adaptation, also narrows the job. It trains small added sets of numerical values rather than changing all the model's original weights. That lets a developer adapt an existing model without repeating the much larger task of building it from scratch.

A cluster name is not a test configuration

Huawei announced the Atlas 950 SuperPoD on September 18, 2025, specifying a system with 8,192 Ascend neural processing units, its AI accelerators. The fine-tuning claim concerns a 128-card setup carrying that product name. ^[3][1]

Those figures describe different things. The product announcement gives the scale of Huawei's advertised system. It does not establish how the reported test was configured or validate its performance. ^[3]

There is separate documentation showing work to support Kimi K3.1 Pro on Huawei hardware. The vLLM-Ascend deployment guide labels its support "Experimental" and describes deployment and evaluation configurations for Atlas 800I A3 and 950DT products. It says support and performance may evolve, but does not document the reported 128-card LoRA fine-tuning run. ^[2]

The next test is reproducibility

Huawei's reported result makes a concrete proposition worth testing: a specified group of its accelerators can adapt a very large model, and a caching change can improve that work. The value to developers lies in whether they can repeat it.

The next useful step is a technical account from Huawei defining what "performance" measures, identifying the comparison configuration and publishing enough detail to reproduce the run. That means specifying the model checkpoint, training data, LoRA settings, numerical precision and training duration.