Xiaomi's MiMo-V2.6-Pro Is Now the Top Open-Weights AI Model

Xiaomi, the company known for phones and rice cookers, just gave away the best open-weights AI model in the world. MiMo-V2.6-Pro, open-sourced September 22, 2026 under the MIT license, tops Artificial Analysis's Intelligence Index among open-weights models, at a fraction of what frontier closed models charge to run.

• MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index v4.3.2, the highest mark of any open-weights model, ranked first of 114 models in its class, ahead of Z AI's GLM-5.3 (max) at 45 and Kimi K3 (max) at 44.
• The model has 1 trillion total parameters but only 42 billion active per token, a mixture-of-experts design behind Xiaomi's $0.43 per million input token and $0.87 per million output token pricing, with a 99% discount on cached tokens.
• Weights for the full Pro model, a lighter Flash variant and a 9B distillation are on Hugging Face under the MIT license, the most permissive open license there is, with commercial use explicitly allowed and no usage caps.
• Anthropic released Claude Opus 5.5 the same day at $4 input and $20 output per million tokens, so September 22, 2026 produced both a cheaper frontier closed model and a top open-weights model priced at roughly a tenth of it.

How does a trillion-parameter model cost 43 cents a million tokens?

The number that should stop you is 1 trillion total parameters against 42 billion active ones. MiMo-V2.6-Pro is a mixture-of-experts model: it holds many specialized sub-networks but routes each token through only a slice, 42 billion parameters' worth. Serving cost follows the active count, not the total, and that's the entire trick behind $0.43 per million input tokens. Xiaomi isn't running a trillion-parameter forward pass on every request, it's running something closer in compute terms to a 42B dense model that draws on a much larger pool of specialization when it needs to.

RelatedLaguna S 2.1: 8B Active Params, 70% Terminal-Bench

The numbers back this up. Output runs at 76.2 tokens per second with a 2.55 second time to first token, fine for interactive apps. Artificial Analysis's full ten-evaluation Intelligence Index run against it generated 140 million output tokens and cost $206.66 total, cheap enough for individual developers to rerun out of curiosity.

What does MIT licensing actually let you do?

Plenty of "open" AI releases aren't really open. They ship under bespoke licenses with revenue caps, field-of-use restrictions, or a separate commercial agreement once you cross some user threshold. MiMo-V2.6-Pro skips all that. It's MIT, full stop: use it, modify it, ship it in a commercial product, no cap, no negotiation. Xiaomi published weights for Pro, Flash and a 9B distillation on Hugging Face under that license, alongside a public technical report.

The model is natively omnimodal on input: text, image, speech and video in, text out, reasoning through extended thinking before it answers. A 1 million token context window holds entire codebases or long documents in one pass. Beyond Pro and Flash there's Pro-UltraSpeed, trading a little latency for up to 20 times faster output at the same quality.

Who is this actually for?

Here's the caveat most coverage will skip. An Intelligence Index score of 46 is a composite across ten evaluations, including AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode and Humanity's Last Exam. A single number compresses a lot of variance, and "top open-weights" is a qualified claim, not "top overall." Frontier closed models remain a separate category; this result says MiMo-V2.6-Pro leads the open field, not that it has closed the gap with every proprietary system.

There's a second, more practical catch. MIT weights are free to download, but a trillion-parameter model still needs serious hardware to self-host, enough accelerator memory for the full parameter set even though only a slice activates per token. Most teams don't have that idle, so for nearly everyone the practical path is a hosted API, Xiaomi's own platform or OpenRouter, not a home GPU cluster.

The same day, a closed frontier model got cheaper too

Anthropic shipped Claude Opus 5.5 on that same news cycle, at $4 per million input tokens and $20 per million output tokens, a 40% cut from Opus 5, scoring 66.4% on Terminal-Bench 4.0. Two companies, two strategies, one day: Xiaomi opened its weights and undercut on price by roughly a factor of ten, Anthropic kept its model closed and cut its own price instead. More on those numbers in our companion piece, Claude Opus 5.5's Terminal-Bench and agentic benchmark results.

RelatedStepFun Step 5 Preview: a 600B Agent Model at $1 per Million Tokens

What it means for the market

This is factual analysis, not investment advice, and none of it should be read as a signal to buy or sell anything. Xiaomi trades publicly on the Hong Kong exchange under ticker 1810, and a publicly traded hardware company just gave away its most capable AI model free, under a license with no strings attached. That only makes sense if the monetization thesis sits outside per-token API margin, most plausibly in devices and ecosystem lock-in, where a strong in-house model makes Xiaomi phones and appliances more valuable without inference itself needing to turn a profit.

The signal for investors to watch is pricing pressure across the stack. A company that doesn't need API revenue pricing a model this capable at $0.43 per million input tokens compresses the ceiling every closed-model vendor and reseller can charge, especially in the thin-margin mid-tier. Watch whether other hardware companies follow suit, whether resellers see pricing power erode, and whether Xiaomi's device revenue shows any lift tied to MiMo adoption.

Our take

The part of this story worth sitting with isn't the leaderboard position, those shuffle every few weeks. It's who did it. A phone and appliance maker just showed that mixture-of-experts architecture has matured enough that giving away a trillion-parameter model no longer requires an AI lab's balance sheet, only a willingness to treat the model as a loss leader for hardware sales rather than a product in its own right. That's a different economic logic than OpenAI, Anthropic or Google run, and pure-play AI labs can't easily match it without a phone business to subsidize the inference bill. Cheap and capable is good news for developers regardless of motive; whether it's good news for the AI lab business model is a separate question, and September 22, 2026 gave a fairly clear early answer.

Original analysis by GenZTech, based on Xiaomi's published model release and independent Artificial Analysis benchmarking.