MiMo V2.6 explained in 12min..

Xiaomi released their new model MiMo V2.6 and their RL environment is quite beautiful. China keeps making their AI advancements open for us to learn and in this case how reinforcement learning can be scaled to be scaled to 10,000+ GPUs to run post-training in just 5 days. Let's learn how they ran asynchronous training on their massive scale of post-training spending huge tokens during the training run. #xiaomi #machinelearning #llm Follow me: X: https://x.com/calebfoundry LinkedIn: https://www.linkedin.com/in/calebeom/ TikTok: https://www.tiktok.com/@calebwritescode Chapters 00:00 Intro 00:38 RL training 03:22 Stale Policy 04:58 Training run 06:22 Sampling 07:11 Token Cost 07:57 Loss Calculation 11:23 RL Infra 12:11 Conclusion