Xiaomi's MiMo v2.6 Takes the Top Open-Weights Spot on the Intelligence Index

A phone company now builds one of the world’s best open AI models. Xiaomi released MiMo v2.6 on September 21, and within hours it debuted at the top of Artificial Analysis’ Intelligence Index for open-weights models: a score of 46, ranked first out of 114 open models, in a size class where the median sits at 18. The specs are frontier-class in every direction at once: a mixture-of-experts architecture with 1.0 trillion total parameters and 42 billion active per token, a 1 million token context window, multimodal input (text, image, speech, and video), and an MIT license that permits unrestricted commercial use.

What Xiaomi Actually Shipped

The architecture choices target exactly the workloads the industry is converging on. The 42B active parameter count keeps inference costs in the same band as far smaller dense models while the 1T total parameter pool preserves reasoning depth, the standard frontier-MoE trade now validated at scale by DeepSeek, Kimi, and Xiaomi alike. The 1M context window is aimed at agentic and codebase-scale work rather than chat. Multimodal input including speech and video is the unusual part: most open-weights reasoning models still ship text-only or text-plus-image, and shipping all four input modes in one checkpoint removes an integration layer every competitor has left to the ecosystem.

The Numbers Behind the Ranking

Artificial Analysis ran the full independent evaluation: 124.5 output tokens per second (rank 12 of 114, well above the 74.8 median), 2.34 seconds to first token (dead average), and an Intelligence Index of 46 against an 18 median for the class. Two caveats belong next to those numbers. First, the index measures a broad capability band, not your workload; a top ranking is a shortlist credential, not a deployment decision. Second, the evaluation notes MiMo is somewhat verbose, generating around the median token count (140M tokens across the suite), which matters because verbosity is billable: the API price is $0.435 per million input and $0.87 per million output tokens, with a 99% cache discount that brings the blended rate to roughly $0.18 per million. Even at full list price those numbers undercut most open-weights rivals’ hosted endpoints, and self-hosting changes the math further.

The MIT License Is the Real Headline

Open-weights leadership has been a Chinese specialty for over a year, but the licenses have been tightening: Qwen’s image models shipped research-only terms days ago, and several Chinese frontier releases restrict commercial use pending separate agreements. MiMo v2.6 ships under MIT, which means any company can take the weights, fine-tune them, run them in production, and ship the result without a negotiation. Chinese open models have been overtaking US ones on Hugging Face downloads for months; an MIT-licensed trillion-parameter model at the top of the capability rankings converts that download momentum into something enterprises can legally standardize on. That is a different competitive act than another high-score release.

What to Watch

Three things determine whether this matters in six months. First, durability: whether independent follow-ups reproduce the Intelligence Index standing, and whether Xiaomi sustains a release cadence (v2.6 implies a v2.x line with real maintenance behind it). Second, distribution: Xiaomi’s own API is currently the sole listed provider, and the model’s practical reach depends on the Groq-class inference shops and hyperscalers picking it up, as they did for Kimi and DeepSeek. Third, the license race: if MIT-licensed frontier models become the norm rather than the exception, the comparison between Chinese open-weights and US closed APIs stops being about capability and starts being about total cost, and that is a argument the closed labs are currently priced to lose.