Xiaomi's MiMo-V2.6 series tops open-weight rankings with security-trained models; Anthropic alleges 400,000+ Claude distillation exchanges

Xiaomi released MiMo-V2.6-Pro-RL (1.02T/42B-active MoE), MiMo-V2.6-Flash-RL (309B/15B), and a 9B distill checkpoint, with Pro ranked first on Artificial Analysis's open-weights index — while Anthropic's threat report GTG-16008 accuses Xiaomi of routing MiMo…

On 2026-09-21 Xiaomi released the MiMo-V2.6 series on Hugging Face: the flagship MiMo-V2.6-Pro-RL (1.02 trillion total parameters, 42 billion activated, 1M-token context, omnimodal text/image/video/audio), the efficiency checkpoint MiMo-V2.6-Flash-RL (309B total, 15B activated, 1M context), and MiMo-V2.6-Distill-Qwen-9B, a 9B agentic checkpoint covering coding, general agent tasks, visual coding, and cybersecurity. Artificial Analysis ranks MiMo-V2.6-Pro first among open-weights models on its Intelligence Index with a score of 46, ahead of Kimi K3 and Qwen, with cited pricing of $0.435/M input and $0.87/M output tokens under an MIT license; Xiaomi also announced a Pro-UltraSpeed mode claiming up to 20x faster generation at the same quality. Xiaomi attributes gains to one mixed reinforcement-learning run spanning coding, agents, vision, and cybersecurity (asynchronous GRPO, 1,568 prompts and 16 rollouts per step on the Flash card), cited at roughly 130 hours / under six days, 75 billion tokens, and $2.6M–$2.62M (Latent Space's headline rounds this to $3M). Sources disagree slightly on DeepSWE: The Decoder says the RL run lifted DeepSWE from 58.4 to 72.6, while Xiaomi's Hugging Face card lists 71.9 for Pro (Flash: 67.9). On security benchmarks, Xiaomi reports 80.2 on MiMo Cyber Bench and 89.9 on Terminal Bench 2.1 for Pro, and 95.1 on CyberGym for Flash; on ExploitBench Pro scores 47.9, below the 78.5 Xiaomi lists for GPT-5.6 Sol. Separately, Anthropic's threat report (case GTG-16008) tracks over 400,000 exchanges routing MiMo conversations through OpenClaw and OpenCode to Claude for what Anthropic calls illegal distillation. Release of training artifacts is partial: The Decoder says Xiaomi open-sourced its RL toolkit with roughly 7,000 auto-graded tasks including OSS-Fuzz cybersecurity tasks, while Latent Space reports the full 7,000-plus task datasets are not yet public and environment code is only promised.

• Three checkpoints released 2026-09-21: MiMo-V2.6-Pro-RL (1.02T total / 42B active parameters), MiMo-V2.6-Flash-RL (309B total / 15B active), and MiMo-V2.6-Distill-Qwen-9B (9B)
• Pro and Flash have 1M-token context and native text, image, video, and audio support
• Artificial Analysis ranks MiMo-V2.6-Pro first among open-weights models on its Intelligence Index with score 46, ahead of Kimi K3 and Qwen
• Cited pricing: $0.435 per million input tokens and $0.87 per million output tokens, under an MIT license
• Mixed RL run cited at ~130 hours (under six days), 75 billion tokens, and $2.6M–$2.62M; Latent Space's headline rounds this to $3M
• DeepSWE figures conflict: The Decoder reports a lift from 58.4 to 72.6, while Xiaomi's model card lists 71.9 for Pro (Flash: 67.9 on DeepSWE v1.1)
• Security benchmarks reported by Xiaomi: Pro scores 80.2 on MiMo Cyber Bench, 89.9 on Terminal Bench 2.1, and 47.9 on ExploitBench (vs 78.5 listed for GPT-5.6 Sol); Flash scores 95.1 on CyberGym
• Anthropic threat report case GTG-16008 tracks 400,000+ exchanges routing MiMo conversations through OpenClaw and OpenCode to Claude, alleging illegal distillation