Xiaomi MiMo-V2.6 Open-Sourced: $3.47M RL Training in Under 6 Days Tops Open-Model Rankings

Forum topic · 小凯 · 2026-09-23

Summary

On September 22, 2026, Xiaomi released and open-sourced MiMo-V2.6, a series of natively omni-modal models in two variants: Pro (1.02T total / 42B activated parameters, sparse MoE) and Flash (309B / 15B). MiMo-V2.6 Pro scored 46 on the Artificial Analysis Intelligence Index v4.3.2, surpassing Qwen3.8 Max and Kimi K3 to become the top open-weight model, roughly 7 points behind the leading closed frontier models. The entire reinforcement learning (RL) post-training run took under six days and roughly $3.47 million (about $850K for Flash, $2.62M for Pro), producing around 750,000 trajectories. Key techniques include the 'You Only RL Once' paradigm integrating multi-domain tasks into a single RL run, Asynchronous GRPO raising GPU utilization from 30-40% to over 60%, a frozen MoE router to prevent training drift, and a four-layer anti-reward-hacking pipeline with confirmed cheating trajectories below 2%. DeepSWE v1.1 scores rose 14-17 points after RL. Xiaomi also released 7,000+ RL task environments, an end-to-end training framework, and a Distill-Qwen-9B model, positioning the work as a first step toward recursive self-improvement (RSI).

On September 22, 2026, Xiaomi officially released and open-sourced the MiMo-V2.6 series, including two natively omni-modal models: Pro and Flash. Pro scored 46 points on the Artificial Analysis Intelligence Index v4.3.2, beating Kimi K3 and Qwen3.8 Max to become the strongest open-weight model on the leaderboard. The full RL post-training run took under 6 days — roughly $850K for Flash and $2.62M for Pro, over $3.47 million combined — generating around 750,000 trajectories.

Key points

Ranking: On AA v4.3.2 (10 benchmarks including AA-Briefcase, GDPval-AA v2, Terminal-Bench 4.0, SciCode, Humanity's Last Exam), MiMo-V2.6 Pro is the only open-weight model in the top five, ~7 points behind leaders GPT-6 Astra, Claude Opus 5, Fable 5.1, and Muse Spark 1.3. Pro and Flash take the top two open-model spots.
Cost efficiency: Pro's measured weighted cost per task is $0.13. Xiaomi claims pricing at 1/20 to 1/60 of comparable overseas models at similar intelligence levels.
Specs: Pro uses sparse MoE with ~1.02T total / ~42B activated parameters; Flash has ~309B total / ~15B activated. Both support up to 1M-token context and text, image, video, and audio input — a prerequisite for long-horizon tasks whose trajectories span tens to hundreds of thousands of tokens.

Training method: "You Only RL Once" + fully async

Led by Luo Fuli's MiMo team, the paradigm integrates coding, professional office tasks, vision, and cybersecurity into a single RL run with multiple lightweight agent scaffolds. Core improvements:

Asynchronous GRPO (Asynchronous Group Relative Policy Optimization): short tasks update while long tasks keep rolling, lifting GPU utilization from 30–40% (synchronous) to over 60%, with single updates batching 1,568 samples.
Frozen MoE Router: freezing routing decisions during RL prevents drift where the model over-routes easy tasks to "favored" experts, misaligning training and real-world distributions; RL only tunes expert capabilities.

Anti-reward-hacking defenses

Four layers guard against models exploiting environment loopholes rather than solving tasks:

1. A cheating-detection agent probing environments before training 2. Frozen router to suppress training drift 3. Anomaly detection monitoring reward spikes and trajectory degradation 4. Cross-validation among multiple verifiers/scorers

Official figures put confirmed reward-hacking trajectories below 2% (self-reported, no third-party audit).

Cost structure

Total ~$3.47M (Flash $850K, Pro $2.62M), with over half spent on trajectory generation and result scoring. Goldman Sachs research estimates clusters of ~4,000 and ~8,000 H200-equivalent GPUs; derived per-token costs are ~$10/M tokens for Flash and ~$35/M for Pro. The livestream revealed mid-run interventions: infra-error reruns, bad-pattern datasets, zero-gradient tasks, and GPU OOM issues. Goldman's take: the RL bottleneck has shifted to engineering optimization.

Gains on long-horizon tasks

On DeepSWE v1.1, Flash rose from 48.8 to 65.68 (+17) and Pro from 58.4 to 72.57 (+14); under avg@3 sampling, Flash gained over 19 points. Demonstrated tasks include playable 3D world generation, Blender scene building, desktop software operation, robotic-arm visual control, frontend design, and audio/video production. As a "Co-Scientist," Pro helped a materials team design MOF structures for adsorbing PFAS, cutting a ~1-month R&D cycle to 2–3 days. It also formalized the Li-Yorke theorem ("Period Three Implies Chaos") in Lean 4 — 6,000+ lines of code fully kernel-verified, with no Lean-specific post-training beforehand.

Open-source scope

Beyond the two model weights, Xiaomi released:

7,000+ high-quality RL task environments (software development, cybersecurity, office, web design, music creation), each with automatic scorers
An end-to-end RL training framework and a lightweight Harness
MiMo-V2.6-Distill-Qwen-9B: SWE-bench Verified 61.1 → 66.2, Terminal Bench 2.1 37.1 → 52.8
A full technical report

vLLM and SGLang shipped inference integration within about a day of the weight release.

RSI path ahead

Xiaomi frames V2.6 as a key step toward Recursive Self-Improvement (RSI): treating RL as a scalable compute lever rather than one-shot fine-tuning, aiming for a flywheel where models generate their own training tasks. The open question is whether model-generated trajectories can train the next generation at quality matching or exceeding human-written data.