Xiaomi MiMo 2.6: What the Benchmarks Actually Show

Hello friends! If I had told you two years ago that the company that makes your vacuum cleaner would ship one of the strongest AI models in the world, you would have laughed at me. Well… the Xiaomi MiMo 2.6 landed this week and did exactly that. And the thing that made me drop everything and write was its price.

Let me explain with a comparison I like. Most AI companies sell you the finished cake: you pay by the slice, you eat it, done. Xiaomi handed over the recipe. The model weights are on Hugging Face under an MIT license, and anyone can download them, run them on their own server, and even sell cake with them. Hold on to that idea, I come back to it at the end.

1. What MiMo 2.6 is (and why the name is confusing)

First, a warning so you do not get lost searching for it. The official name is MiMo-V2.6, with a “V”, and it is not one model: there are three.

MiMo-V2.6-Pro is the flagship, with 1.02 trillion total parameters and 42 billion active per token. MiMo-V2.6-Flash is the budget version, with 309 billion total and 15 billion active. And MiMo-V2.6-Pro-UltraSpeed is the same Pro, only up to 20 times faster at generating, charging ten times more for it.

All three are omnimodal, meaning they understand text, images, video and audio. And all of them have a 1 million token context window, something like 1,600 pages of text at once.

2. The benchmarks: how far MiMo 2.6 gets

The number going around is from the Artificial Analysis Intelligence Index, an independent index that combines ten different evaluations across math, science, coding and reasoning. MiMo-V2.6-Pro scored 46 and became the highest-placed open-weight model in the world, tied with Grok 4.7 and one point behind GPT-5.6 Sol.

Look at the length of those bars: they are almost the same. That is the news. A model you can download for free pulled up alongside the closed ones that cost a fortune.

On task benchmarks, Xiaomi published its numbers side by side with the closed competitors, on the MiMo-V2.6 model card and in the official announcement. Here I have to be straight with you: the company built that table itself, which calls for some skepticism.

Notice that the scoreboard is split. MiMo wins on Terminal Bench and AutomationBench, sits in the middle on the visual test, and loses on DeepSWE v1.1, the one closest to real programming work: 71.9 against 74.0 for Claude Opus 5. And on competitive programming it does clearly worse. Open and cheap does not mean better at everything.

There is also CyberGym, on security, where Xiaomi reports 94.0 for Pro and 95.1 for Flash. I left it out of the chart because the table carries no competitor numbers for that test, so there would be nothing to compare it against.

3. The price is the part that actually impressed me

MiMo-V2.6-Pro charges $0.87 per million output tokens. Claude Opus 5 charges $25 and GPT-5.6 Sol charges $30. This is not 20% cheaper, it is almost 30 times cheaper. And Flash, at $0.28, basically disappears off the chart.

I have written here before about the AI price war (in Portuguese) and about Jev, from TypeSafe AI, which took the same route: charge almost nothing and leave the competition explaining the difference. MiMo 2.6 is another shove in that direction, and anyone who uses AI daily gains from it.

There is another number that says a lot: training Pro cost around $2.62 million, and Flash around $850,000, in under six days. For a frontier model, that is not much money.

4. How to try MiMo 2.6 today

You do not need a server of your own to experiment:

1. Through the website. Open MiMo AI Studio and talk to the model right in the browser, the way you would with any chatbot.
2. Through OpenRouter. If you already use a tool that accepts OpenRouter, MiMo-V2.6 is there and you switch models in a dropdown.
3. By downloading the weights. On Hugging Face, under an MIT license that allows commercial use. There is also a distilled 9 billion parameter version, much lighter for anyone who wants to run it at home.

Xiaomi also opened more than 7,000 reinforcement learning environments alongside the weights, which is the material other researchers use to train their own models. That is rarer than open sourcing the model itself.

5. What this changes for you

If you are a developer, it is worth testing Flash on high-volume tasks, the ones where today you just swallow the cost of the expensive model because there is no alternative. The price gap is big enough to change what is viable to build.

If you do not write code, the effect reaches you more slowly, but it does reach you: every strong open model that shows up pushes everyone else’s prices down and the quality of the free tier up.

And the skepticism I recommend keeping: a benchmark published by the company itself is a shop window. Xiaomi’s numbers can be perfectly accurate and still not describe your project. Test it on your own case before you switch.

6. Frequently asked questions

Is Xiaomi MiMo 2.6 free?
The weights are, under an MIT license, and you can download and run them yourself without paying Xiaomi anything. Using their API costs money, but it runs at $0.435 per million input tokens and $0.87 output on Pro.

Is MiMo 2.6 better than Claude or GPT?
On some agent and terminal tasks, Xiaomi’s numbers say yes. On heavy coding, no: Claude Opus 5 and GPT-5.6 Sol still do better on DeepSWE. Where it wins comfortably is the ratio between result and price.

Can I use MiMo 2.6 in a commercial product?
You can. The MIT license allows commercial use, modification and redistribution. Just think carefully about where your data ends up if you use the API instead of running the model on your own server.

Back to the cake: this time the recipe came with it, and that is what makes the difference.

Cheers, Fellipe Soares