MiMo V2.6 Flash (Consensus Protocol)

Xiaomi released the MiMo V2.6 family on September 22, 2026, led by the 1.02T-parameter Pro and 309B-parameter Flash sparse-MoE models, both supporting 1M-token context and multimodal input across text, images, video, and audio. MiMo-V2.6-Flash targets high-volume automation and cost-sensitive agent loops with 15B active parameters. Xiaomi open-sourced the Pro and Flash RL checkpoints under the MIT license, alongside supporting environments and training resources. The hosted MiMo-V2.6-Flash API is priced at $0.14 per million input tokens and $0.28 per million output tokens, compared to $0.435 and $0.87 for Pro. The release includes a smaller MIT-licensed Qwen-based distill (MiMo-V2.6-Distill-Qwen-9B) for local experiments and RL research, plus a Pro-UltraSpeed hosted serving mode for low-latency interactive work. Xiaomi recommends Flash for cost-sensitive workloads and Pro for complex engineering and long-running tasks.

Xiaomi's MiMo team shipped MiMo-V2.6 on September 22, 2026, five days after publicly livestreaming the model's reinforcement-learning training run. The launch hit 558 points on Hacker News, with the top comment praising Xiaomi's transparency over benchmark scores. The full technical report, "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement," was posted on Hugging Face Papers and arXiv (2610.11959) by October 9, 2026. The technical report describes the omni-modal family starting from mid-training on a broad multimodal corpus, then scaling RL compute along three axes: bigger batches (1,568 samples, 2.7–3.7 billion tokens per step at up to 1M context), more varied environments (code, general, visual, cyber tasks), and groupwise agentic grading for better long-horizon rewards. The team froze the MoE router and built multi-layer defenses against reward hacking, while open-sourcing training dynamics, RL environments, and the RL framework. On CyberBench v1.1, MiMo-V2.6 Flash is listed first among nine security-bug benchmark models.