AI BriefingNVIDIA Releases Kimi NVFP4
·2026.05.14 21:53
Key point
NVIDIA released NVFP4 versions of Kimi-K2.6 and Kimi-K2.5.
Details
NVIDIA quantized Moonshot AI's Kimi-K2.6 into NVFP4 and released it on Hugging Face. Kimi-K2.5-NVFP4 was also uploaded alongside it.
- The quantized version was created with Model Optimizer.
- It is marked as available for commercial and non-commercial use.
- The baseline is the native INT4 version of Kimi-K2.6, and benchmarks were measured under conditions of
temperature=1.0,top_p=0.95,max num tokens=128000. - The figures are GPQA Diamond 90.9→90.4, SciCode 52.6→54.4, τ²-Bench Telecom 98.2→98.0, MMMU Pro 75.6→76.5, AA-LCR 71.0→71.8, IFBench 73.9→73.9.
In terms of numbers, SciCode, MMMU Pro, and AA-LCR increased, GPQA Diamond and τ²-Bench Telecom decreased slightly, and IFBench stayed the same.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
2026 AI Inference Hardware Innovations: New Architectures for Prefill/Decode Disaggregation and Solving Memory Bottlenecks
Stanford and NVIDIA Release 'Contrastive Language Model' CLM with 9x Faster Inference