Peano AI
An AI team specializing in large-scale model training infrastructure, with expertise in full-parameter reinforcement learning training of very large models on TPU clusters.
Timeline (last 90 days)
The Peano AI team ran full-parameter reinforcement learning for Xiaomi's 310B-parameter MiMo-V2.6 model on a cluster of more than 1,000 TPUs.
The team achieved bitwise-matched consistency between rollout and the trainer during the validation stage, ensuring that model computations during inference and training are numerically identical.
The specific training results, magnitude of model performance improvement, and cost data for this work have not yet been made public.
All Facts (2)
The team achieved bitwise-matched consistency between rollout and the trainer during the validation stage, ensuring that model computations during inference and training are numerically identical.
50%UnverifiedThe Peano AI team ran full-parameter reinforcement learning for Xiaomi's 310B-parameter MiMo-V2.6 model on a cluster of more than 1,000 TPUs.
50%