@tianjun_zhang: We scale up RL on TPUs for the MiMo families It is quite a journey for us to bring RL to scale with Jax + TPU: the most…
Summary
Peano Labs has scaled reinforcement learning on TPUs for the MiMo model family, enabling full-parameter RL at 310B parameters with Jax, where scaling is primarily a configuration change.
View Cached Full Text
Cached at: 09/22/26, 04:04 PM
We scale up RL on TPUs for the MiMo families 🚀
It is quite a journey for us to bring RL to scale with Jax + TPU: the most amazing part for me is that scaling it up is mostly a config change, not a code rewrite. This is the dream for any researcher 🫣
Read our blog for more details. Thanks to @XiaomiMiMo, @berkeley_ai, and @googlecloud for making it happen! We’ll share more details soon!
Peano Labs (@peano_ai): We enable full-parameter RL on TPUs: MiMo-V2.6 at 310B, plus other stable training runs of 1,000+ steps across 1,000+ TPUs.
With JAX, scaling up is a config change, not a rewrite.
We built on that with optimized vLLM inference for faster rollouts and full bitwise
Similar Articles
@_LuoFuli: MiMo-V2.6: The Hard Road to Scaling Up RL MiMo-V2.6 is very likely one of the largest single RL runs, by compute, that …
MiMo-V2.6 is a large-scale reinforcement learning model that has become the top open-source model, with research innovations surpassing DeepSeek R1 and resources released to advance Agentic RL research.
@ying11231: Impressive performance on TPU.
A blog post from LMSYS Org details optimizing Ling-2.6-1T, a 1 trillion parameter hybrid MoE model, on TPU v7x using SGLang-JAX, achieving efficient inference by hiding MoE data movement behind computation with a single Pallas kernel.
@samsja19: prime-rl can now train 1T parameters MoE blazingly fast, under 5 minutes per step, or 1k steps in ~3 days To achieve th…
Prime Intellect released prime-rl v0.6.0, enabling reinforcement learning at trillion-parameter MoE scale with sub-5-minute step times and optimized inference, training, and rollout.
@googledevs: A major update to Tunix for scaling Agentic RL is here The new asynchronous, decoupled rollout engine solves multi-turn…
Google announces a major update to Tunix, its post-training library, with an asynchronous decoupled rollout engine to scale agentic reinforcement learning on JAX/TPU, eliminating idle time and improving throughput.
@chumacn: That's awesome, Luo Fuli is so cute, please Jensen Huang @JensenHuang give us more cards, just a few hundred B300s are …
The tweet provides an update on MiMo-V2.6's reinforcement learning scaling efforts, including compute and environment enhancements, and requests more NVIDIA B300 GPUs for further development.