Xiaomi MiMo-V2.6-Pro: a phone maker's open frontier model
What Xiaomi's MiMo-V2.6-Pro is, what its MIT-licensed open weights allow, the benchmarks Xiaomi reports, and why most businesses would use it hosted.
Xiaomi is best known for phones, but it also trains large language models. In late September 2026 it released MiMo-V2.6-Pro, the flagship of its MiMo-V2.6 series, and published the weights on Hugging Face under the MIT License. Anyone can download the model, run it and modify it.
That makes it a large, capable model that a business can inspect and host itself, within the licence terms. It is also far too large for an office computer. For most businesses, the useful question is not whether it is the best model, but whether you can use it, where it would run, and on what terms.
What MiMo-V2.6-Pro is
MiMo-V2.6-Pro accepts text, images, video and audio as input and returns text. Like DeepSeek’s V4 models, it is a mixture-of-experts model, which means only part of it works on each token. Our post on DeepSeek V4 and V4-Pro explains the idea. The model card gives 1.02 trillion parameters in total, with 42 billion active at each step.
The context window is 1 million tokens. Xiaomi’s card says that window is meant for “long repositories, tool traces, and multi-session agent runs”: long jobs where a model reads code, calls tools and keeps track of many steps of work.
It is the third Pro model Xiaomi has released this year. The first, MiMo-V2-Pro in March, was available only through Xiaomi’s API. MiMo-V2.5-Pro, in late April, was the first with open weights, and V2.6-Pro keeps the same size and the MIT License while adding the wider range of inputs.
Xiaomi also publishes a smaller sibling, MiMo-V2.6-Flash, with 309 billion total parameters, 15 billion active, the same four input types and the same 1 million token window. Its card lists several ways to use the series without downloading anything: Xiaomi’s AI Studio, MiMo Code, Xiaomi MiMo Desktop, the Xiaomi MiMo Open Platform API and OpenRouter.
Benchmarks Xiaomi reports
The model card carries a long table of results. These are Xiaomi’s own numbers, not independent tests, so read them as claims. Four of them show where the company is pointing the model:
- DeepSWE v1.1, a coding agent test: 71.9.
- Terminal Bench 2.1, tasks run in a command line: 89.9.
- OSWorld-Verified, tasks carried out on a desktop computer: 82.0.
- CyberGym, security tasks: 94.0.
Nearly every test in the table is about agents, models that work through multi-step tasks with tools, rather than chat or writing. That tells you what Xiaomi built it for. It does not tell you how the model will handle your invoices or your support inbox. Treat the table as a reason to test the model, not as the result of a test.
What an open release gives you
Xiaomi’s homepage describes the MiMo series as “frontier intelligence, all the modalities, built in public.” The company has not said how it plans to use a model of this size in its own products, and we will not guess.
What matters for a business is what the open release gives you, whatever the reason behind it. Published weights can be tested, hosted and modified by anyone who accepts the licence. You are not tied to one company’s service, pricing or decision to retire a model. If Xiaomi changes its API terms tomorrow, the weights you downloaded still work. Open weight and open source are not the same thing, and it is worth knowing which one a model offers before you plan around it.
What it means for a small business
There are three practical points.
A capable model you can evaluate freely. The MIT License is short and permissive. Read the copy that comes with the weights before a commercial product depends on it, as you would with any licence.
Not something to run on an office PC. At 1.02 trillion parameters, the weights alone run to hundreds of gigabytes even in compressed (quantised) form, and close to a terabyte at 8-bit precision. Running the model yourself means server hardware and someone to manage it, which is a different decision from buying a computer. For teams that do host it, the card points to SGLang, a model-serving tool, for best performance.
Hosted access is the realistic route. Most small businesses would reach the model through an API, where you pay per million tokens processed. Prices are set by each host and change often, so compare the current rate on the platforms you are considering. Check where each host processes your data before you send it anything sensitive.
A first test worth running
Pick three real tasks from your own work, such as summarising a customer email thread, drafting a quote from a standard template or answering questions from your product manual. Run them through a hosted copy of MiMo-V2.6-Pro and through the model you use now, and compare both with what your team would have written. If the open model holds up, that is when it is worth talking about connecting it to your systems as an AI agent. If it does not, you have lost an afternoon, not a budget.
Sources
Start with the people doing the work.
We build software, AI chatbots, and AI agents around the systems you already run, with people keeping the judgment.