Xiaomi releases MiMo-V2.6-Pro-RL with 1M-token context and 1.02T parameters

The architecture uses a Sparse Mixture of Experts with 1.02 T total parameters and 42 B activated parameters. It includes a 681M‑param MiMo ViT vision encoder, a 308M AudioTokenizer plus a 127M audio patch encoder, and a 5‑layer speculative decoder. Training employed Group Relative Policy Optimization on very large batches of 1,568 prompts × 16 rollouts per step, with billions of tokens per update, and a groupwise agentic grading loop that ranks rollouts to drive self‑improvement. Deployment instructions reference SGLang and vLLM Docker images.