Our Blog

Blog Index 

Xiaomi MiMo-V2.6-Pro: A $3 Million Training Run Crowns a New Open-Weights Champion

Posted on 24th Sep 2026 06:03:41 in Artificial Intelligence, Machine Learning

Tagged as: Xiaomi, MiMo-V2.6, Open Source AI, AI Models, AI Benchmarks

Xiaomi has entered the frontier AI race, and it did not arrive quietly. On September 21, the Chinese consumer-electronics giant released the MiMo-V2.6 family: two natively omnimodal open-source models whose flagship now sits at the very top of the open-weights class. MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index (v4.3), passing Z.ai's GLM-5.3, Moonshot's Kimi K3 and Alibaba's Qwen3.8-Max to become the strongest open-weight model available anywhere. The weights are published on Hugging Face under an MIT license.

The surprise is not the score alone. It is who delivered it, and how cheaply. Xiaomi is best known as a smartphone and smart-home manufacturer, never counted among the Chinese AI labs that analysts track. Its run to the top took under six days of reinforcement learning at a cost of roughly $2.6 million for the Pro model, and Xiaomi streamed the entire training run live in public, metrics and all.

What Xiaomi Released

The MiMo-V2.6 series is built around two omnimodal models that accept text, image, speech and video natively and produce text output, plus an ultra-fast serving variant:

  • MiMo-V2.6-Pro — a sparse mixture-of-experts model with 1.02 trillion total parameters, of which only 42 billion activate per inference. It is Xiaomi's most capable model to date and carries a one-million-token context window.
  • MiMo-V2.6-Flash — roughly 310 billion parameters with 15 billion active, positioned as the best balance of intelligence, efficiency and cost.
  • MiMo-V2.6-Pro-UltraSpeed — a serving mode that delivers up to 20 times the output speed at the same quality, aimed at latency-sensitive workflows.

API pricing is unchanged from the outgoing V2.5 generation. MiMo-V2.6-Pro costs $0.435 per million input tokens and $0.87 per million output tokens, while Flash costs $0.14 and $0.28 respectively. Cached input drops to $0.0036 and $0.0028 per million tokens. The models are available through Xiaomi's AI Studio, the MiMo Code environment, the MiMo Desktop app, the company's own API platform and OpenRouter.

How It Ranks Against the Frontier

On Artificial Analysis' Intelligence Index, MiMo-V2.6-Pro lands at 46, first among open-weight models and sixth overall when closed models are included. Every model above it is proprietary: Claude Fable 5.1 and GPT-6 Astra sit at 53 points, Claude Opus 5 at 51, Meta's Muse Spark 1.3 at 48 and GPT-5.6 Sol at 47. xAI's Grok 4.7 ties Xiaomi at 46. Put differently, the best freely downloadable model in the world is now within a single point of the best closed model a developer can buy, and seven points from the absolute frontier.

The cost picture is where Xiaomi presses hardest. Artificial Analysis measures MiMo-V2.6-Pro at roughly $0.13 per Intelligence Index task, placing it on the Pareto frontier of intelligence versus cost: GPT-5.6 Sol-class capability at a fraction of the serving cost. Throughput is measured at about 125 output tokens per second, ranking it 12th of 114 tested models. The benchmarking service does flag one weakness worth noting: verbosity. The model burned 140 million output tokens to complete the index run, which inflates real-world costs for some workloads.

On task-level benchmarks, Xiaomi's published results tell a nuanced story:

  • Cybersecurity: Pro scores 94.0 on CyberGym, and Flash takes first place outright at 95.1, ahead of DeepSeek V4.1 Flash (88.1) and GLM-5.3 (84.5).
  • Workflow automation: Pro reaches 53.1 on Automation Bench, just behind DeepSeek V4.1 Flash (54.8) and ahead of GPT-6 Astra (52.0) and Claude Opus 5 (50.3).
  • Software engineering: Pro hits 71.9 on DeepSWE v1.1, close to DeepSeek V4.1 Flash (74.2) and the 74.0 scored by Claude Opus 5 and GPT-6 Astra.
  • Tool use: 76.9 on Toolathlon-verified, ahead of GPT-5.6 Sol (74.9) and narrowly behind the Claude models.
  • Economically relevant work: 1,673 on Artificial Analysis' GDPVal, third overall behind Claude Fable 5.1 (1,735) and Claude Opus 5 (1,708).

The gaps run the other way on long-horizon terminal work: 34.9 on Terminal Bench 4.0 against 59.6 for GPT-6 Astra and 49.0 for Claude Opus 5, and 17.8 on ExploitGym. Marathon agentic sessions that grind through a terminal for hours remain a frontier-model strength.

Built in Public: The $3 Million RL Run

Xiaomi credits the leap to scaled reinforcement learning on verifiable, complex tasks. In under six days, both models completed 30 RL steps over roughly 750,000 trajectories each. The Pro run cost about $2.62 million and the Flash run about $0.85 million, with the flagship training at roughly 130 hours and 75 billion tokens. The company livestreamed the production run as it happened, an unusual level of transparency that drew attention across the AI community.

The RL stack scaled along three axes: larger batches and higher throughput, with 1,568 samples per update and training at up to a one-million-token context length, consuming 3.5 to 3.7 billion tokens per step; a multi-task training suite spanning coding, general agents, visual and cyber work, mixed across several harnesses so gains in one capability reinforce the others; and more grader compute, using relative comparison within task groups to give long-horizon tasks sharper reward signals.

To keep the run stable, Xiaomi froze the router to suppress training drift and built layered defenses against reward hacking, spanning reward design, adversarial evaluation, anomaly detection and cross-checking between verifiers. The results on held-out benchmarks show the run paid off: on DeepSWE v1.1, Pro climbed from 58.4 to 72.6 points and Flash from 48.8 to 65.7, with scores still improving when the runs ended.

With the weights, Xiaomi is releasing its full technical report, the training environments and the RL code, including recipes for coding, vulnerability reproduction, general knowledge work, visual web development and music generation. The complete 7,000-plus task datasets have not yet been released. One caveat for the headline number: the roughly $3 million figure covers the reinforcement-learning phase, not pretraining of the base model the run started from, which already scored in the high 40s on DeepSWE v1.1 before any of it.

Why It Matters

MiMo-V2.6-Pro extends a cadence that has become hard to dismiss. In roughly ten weeks, Chinese labs have shipped Kimi K3, Qwen3.8-Max, DeepSeek V4-Pro, GLM-5.3, the Tencent Hy4 preview and more, and Xiaomi now pushes the open-weights ceiling past all of them. For developers, startups and enterprises that want to self-host or fine-tune, the answer to what is achievable without license fees and without tying workloads to a US provider has moved again.

The more interesting shift is what Xiaomi is giving away beyond the weights. Researchers argue that high-quality open RL environments, the harnesses that generate and grade training tasks, may now be as strategically important as pretraining corpora were in the previous cycle. Xiaomi is publishing that tooling, which makes the run reproducible in principle, not just the artifact downloadable.

The model's own capability demos point in the same direction. Xiaomi pitches a move from vibe coding to what it calls vibe world: generating 3D scenes with interaction logic from text or video prompts, modeling objects in Blender, and controlling a Franka Panda robotic arm in simulation through multiple camera feeds. In research collaborations, the model designed a metal-organic framework for binding PFAS pollutants using open-source simulation tools, and helped formalize the main theorem of the 1975 paper Period Three Implies Chaos in Lean 4, a proof spanning more than 6,000 lines that the kernel verified in full, despite no Lean-specific post-training.

None of this settles the debate over whether closed models retain a durable edge, and practitioners on Hacker News were quick to note that the eye-catching training cost covers post-training, not the base model behind it. But the direction is consistent: the gap between the best open and closed systems is now measured in single-digit index points and cost multiples, not years.

What to Watch

Three things will decide how far this lands. First, independent evaluation: several benchmark scores in Xiaomi's tables are self-reported, and third parties have not yet reproduced the full picture. Second, the release of the promised RL environments and task datasets, which would give other labs a genuine reproduction path. Third, adoption pressure: if enterprises begin routing production workloads to a 46-point model that costs a tenth of the alternatives, the pricing conversation across the industry changes with them. For now, the open-weights leaderboard has a new name at the top, and it is not one the incumbents saw coming.

Sources

Recent Posts

Catgories

Archives

whatsapp me