Xiaomi has significantly expanded its AI activity with MiMo-V2.6, a new model family including MiMo-V2.6-Pro and MiMo-V2.6-Flash.
Unlike an announcement limited to a demo, Xiaomi also published weights, an API, a technical report and reinforcement learning materials, and the models are available through AI Studio, MiMo Code, MiMo Desktop and OpenRouter.
The company presents support for text, images, video and audio, and a context window of up to one million tokens.
The launch is particularly interesting because it combines three things that do not always come together: large models, relatively low API prices and public access to a significant share of training resources.
Pro, Flash and UltraSpeed
Xiaomi offers several usage options.
MiMo-V2.6-Flash is the cheaper model. Uncached input pricing starts at 0.14 dollars per million tokens, and output at 0.28 dollars.
MiMo-V2.6-Pro costs 0.435 dollars per million input tokens and 0.87 dollars for output.
There is also an UltraSpeed option, with significantly higher prices of 4.35 dollars for input and 8.70 dollars for output per million tokens.
Xiaomi claims UltraSpeed mode can deliver up to 20 times the output speed while maintaining the same quality.
This is a company claim that still requires independent testing across different use cases.
What was actually made public
One interesting aspect of the launch is that the company did not stop at providing an API.
The MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL weights were published on Hugging Face, alongside a technical report and reinforcement learning resources.
The materials are offered under the MIT license.
According to the details published on Hugging Face, Pro-RL is a 524 billion parameter model, and Flash-RL has 159 billion.
There is also MiMo-V2.6-Distill-Qwen-9B, a much smaller model intended for agentic scenarios and more accessible in terms of hardware requirements.
"Open" does not mean everything can run on a laptop
The availability of model weights does not make a model with hundreds of billions of parameters easy to run.
Models such as 159B and 524B require substantial infrastructure, especially when good performance and reasonable response times are needed.
For most users and businesses, the API will therefore probably be a simpler route than self-hosting the large models.
The 9B version, on the other hand, may be more interesting for those who want to experiment with local deployment or limited hardware.
What you can do with MiMo-V2.6
Xiaomi presents the series as a multimodal model family that can work with text, images, video and audio.
The company also presents scenarios involving computer use, content creation, three-dimensional tasks, music, video and research.
But a long list of capabilities does not mean they are all at the same level.
Anyone considering the model for a business use should test their specific task.
A model that excels at code is not necessarily the best for image analysis, and a model with a large context window does not necessarily use all that context equally efficiently in every task.
One million tokens matters, but it is not the whole story
Xiaomi states that the context window supports up to one million tokens.
This can be useful for analyzing code repositories, large documents or processes where a great deal of information needs to be included in the same conversation.
But context window size alone does not guarantee that the model will maintain the same accuracy throughout the text.
Costs can also change significantly when working with such token volumes.
It is therefore important to test not only whether the model "supports a million," but what happens to quality, time and cost when a substantial part of the window is actually used.
What the benchmarks say
Xiaomi presents Pro as one of the strongest open models on the benchmarks it relies on.
According to the company, it scored 46.32 on Artificial Analysis Index v4.3.
The company also published data on RL training costs and time.
All these figures are interesting, but there is still not enough independent testing to determine how MiMo-V2.6 performs in the real world against a broad range of models.
The conditions of each comparison also need to be checked: tool use, reasoning level, number of attempts and other settings can change the result.
What about Hebrew?
For now, the reporting contains no verification that would establish this as a particularly strong model in Hebrew.
The fact that a multilingual model can accept Hebrew text does not mean its quality is the same as in English or Chinese.
Israeli companies considering using it should test writing, instruction following, business documents, Israeli names and mixed Hebrew and English themselves.
Why the launch is interesting
MiMo-V2.6 illustrates how competition in the model market is no longer limited to OpenAI, Anthropic and Google.
Chinese companies offer large models, aggressively priced APIs and open weights that let developers choose between a cloud service and self-hosting.
For anyone evaluating a new model, the question is not whether Xiaomi published an impressive number in a table.
The question is whether the model delivers quality, stability and cost suited to the actual workflow.