Biggest open RL run yet: 2,048 GPUs, 9 days, 17M rollouts, 3T+ tokens
LiTianleli · x · 2026-11-06
Details of the Inkling joint RL run suggest the largest open RL training run with public metrics: 1,075 steps, 17M+ rollouts over 9 days on 2,048 GPUs, over 4×10²³ FLOPs and 3T+ tokens. Batch size (3B tokens) matches MiMo V2.6 but with 36x more steps; unlike the previous Inkling release, reasoning and agentic RL were combined in a single joint run instead of separate runs merged via multi-teacher distillation.
More from Infra
- SGLang Ships Day-0 Kimi K3.1 Support With Out-of-the-Box Megakernel — BanghuaZ · 2026-11-06
- SK Hynix says DRAM and HBM volumes through 2029 are effectively sold out — Beth_Kindig · 2026-11-06
- All 22 DeepSeek V4.1 Flash endpoints match reference — 3.9-point range, most consistent model in the Index — ArtificialAnlys · 2026-11-06
- GLM-5.3 has the most endpoints in the Accuracy Index but widest spread: 67.2 to 99.3 across 26 endpoints — ArtificialAnlys · 2026-11-06
- Artificial Analysis launches Endpoint Accuracy Index v1.1: nearly 1 in 4 serverless endpoints score below reference — ArtificialAnlys · 2026-11-06
- Zyphra ships ZYFA: FP8 attention on AMD MI355X at 1.16-1.76x AITER throughput — bclavie · 2026-11-06