Miles achieves end-to-end MXFP8 and per-token NVFP4 reinforcement learning on Blackwell architecture
Original title:Blog Towards Blackwell-Native 8-bit and 4-bit RL: End-to-End MXFP8 and NVFP4 RL in Miles TL;DR: We implemented two Blackwell-native RL recipes in Miles: end-to-end MXFP8 and per-token NVFP4 for MoE experts. Both are supported by fine-grained precision control across checkpoint conversion,... Ziang Li, humans& and Miles Team
AI summary
The Miles team has implemented two native low-precision reinforcement learning schemes on the Blackwell architecture: end-to-end MXFP8 and per-token NVFP4 for MoE expert weights. In ablation experiments on Qwen3-30B-A3B with 8x B200, the raw reward curves of BF16 and all five low-precision configurations closely overlapped, and MXFP8 and NVFP4 reduced inference time.
Key Takeaways
- The Miles team has implemented two native low-precision reinforcement learning schemes on the…
- In ablation experiments on Qwen3-30B-A3B with 8x B200,the raw reward curves of BF16 and all five…
- and MXFP8 and NVFP4 reduced inference time.
Quick Facts
- Published
- Updated
- Source
- LMSYS Blog
- Category
- Papers & Research
- AI score
- 69
虽然只针对Blackwell,但这份报告把MXFP8和NVFP4在RL里的实现细节讲清楚了,从量化合约到每token缩放都给了可复现配置,做大规模RL训练的能直接上手。
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Miles achieves end-to-end MXFP8 and per-token NVFP4 reinforcement learning on Blackwell architecture. July 29, 2026. https://new.nat6.net/en/news/cms6drkj8001crohzuxf7kfy7/