Announcement
SkyRL adds day-zero reinforcement learning support for Reflection’s Beam
SkyRL says the integration supports LoRA, BF16 full fine-tuning and MXFP8 reinforcement learning out of the box.
TLDR
SkyRL says it worked with Reflection and Inferact to add day-zero reinforcement learning support for Beam, a 501B-parameter mixture-of-experts model. SkyRL reports that AIME25 avg@8 accuracy rose from 61% to 79% over 50 training steps. It also reports a 44% throughput improvement on a simple math task using MXFP8, with performance at parity with BF16 training.
Combined views
2.2K
3 Sources, first seen ago
