Skip to content
View original post on X: LMSYS OrgOfficial
Published
34/100

miniTE trains learner updates 1.33–1.37× faster than BF16 on 8×B300

AIInfiX.ai's miniTE, a low-precision training framework, runs learner updates 1.33–1.37× faster than BF16 on 8×B300 GPUs, according to LMSYS Org. The framework diagnoses training–sampling mismatch from quantization to token probabilities and uses TRIAGE to control risky updates.
LMSYS OrgVerified on XOfficial
@lmsysorg
View on X

🚀 miniTE: 1.33–1.37× BF16 speed for learner updates on 8×B300.

InfiX.‌ai’s miniTE is a low-precision training framework that reduces training–sampling mismatch and speeds up NVFP4 computation.

Highlights:
- Diagnose mismatch from quantization to token probabilities.
- Control risky updates with TRIAGE.
- Reduce compute overhead with grouped GEMM, contiguous buffers, and fusion.

Read the full report: http://lmsys.org/blog/2026-10-25-minite

Source details

LMSYS Org · x.comPublished · added here

Original title: 🚀 miniTE: 1.33–1.37× BF16 speed for learner updates on 8×B300.

Topics

Deployment & engineeringData & trainingProduct updates