The Top-Ranked Open AI Model Just Became Free

A small AI research team just put the world's best open-weight language model out for free.

What's happening

On September 22, Xiaomi's AI lab released MiMo-V2.6, an MIT-licensed model that landed at number one on the Artificial Analysis Index, beating all 114 other tracked open-weight models. It cost about $3 million to train, which is a rounding error for a big tech company and roughly what a mid-size firm spends on one vendor contract.

There are two versions worth knowing about. The Pro version scores highest on benchmarks. The Flash version is the one a small team should care about: it costs $0.28 per million output tokens through the API, compared to $5 to $15 per million for most proprietary models. That is a 20x difference in price for a model that outranks all the open competition. There is also a 9B distillation you can run offline on a decent laptop through Ollama, no API key required.

This matters because the capability gap between "what big companies run" and "what a three-person team can afford" keeps closing. Each time a genuinely capable model goes free under an MIT license, the old excuse that quality AI costs a lot gets harder to defend.

Try this this week

• If you are already paying for an AI writing or summarizing tool, check whether it lets you swap the underlying model. OpenRouter (openrouter.ai) connects to MiMo-V2.6-Flash and lets you compare costs side by side.
• If Ollama is installed on your machine, run ollama pull mimo-v2.6-9b and test it on a real task you use AI for today: summarizing meeting notes, drafting a client email, pulling key points from a document.
• Run a quick cost comparison. Take one task you run regularly with a paid AI tool and estimate how many times a month you do it. Plug that into the MiMo Flash pricing. The number is often surprising.

The bigger picture

The trajectory is consistent: frontier-grade AI goes open source, prices fall, and small orgs gain access to tools that used to require a large vendor contract. The build9 approach has always been to find the smallest working thing you can run yourself, and the smallest working thing keeps getting more capable. MiMo-V2.6 is a good reminder to revisit the AI subscriptions you are paying for and ask whether a free alternative has caught up.