Science & Tech · AI · 16 hrs ago
Huawei Ascend tests LoRA fine-tuning of Kimi K3.1 Pro on 128 cards
Ascend says it validated LoRA fine-tuning of the Kimi K3.1 Pro model.
The test used a 128-card Atlas 950 SuperPoD, a computing system built around 128 cards.
Kimi K3.1 Pro has 2.8 trillion parameters, the adjustable values that help a model produce its outputs.
LoRA is a way to fine-tune a model without updating all of its parameters.
Ascend says Swap Core caching improved performance by 15% on average in the test.
This suggests the caching method can help when fine-tuning a very large model, though no further details or independent comparison are given.
The account does not say what happens next.
Huawei Ascend said it validated LoRA fine-tuning of Kimi K3.1 Pro on a 128-card Atlas 950 SuperPoD.
The model has 2.8 trillion parameters, according to Pandaily.
Ascend cited an average gain of 15% from Swap Core caching.
Pandaily published the report on November 8, 2026.
- Who
- Huawei Ascend.
- What
- It said it validated LoRA fine-tuning of Kimi K3.1 Pro on a 128-card Atlas 950 SuperPoD.
- When
- November 8, 2026, the date Pandaily published the report.
- Where
- On a 128-card Atlas 950 SuperPoD. The sources do not state a physical location.
- Why
- Not stated.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- Model
- Kimi K3.1 Pro
- Parameters
- 2.8 trillion
- System
- 128-card Atlas 950 SuperPoD
- Reported gain
- 15% average from Swap Core caching
- Report date
- November 8, 2026