What Is MiMo-V2.6? Performance and Pricing of Xiaomi's Open-Weight AI | CoAI

On September 22, 2026, China's Xiaomi, known for its smartphones, home appliances and electric vehicles, announced the MiMo-V2.6 series of AI models. Centered on two models, the higher-end MiMo-V2.6-Pro and the lighter, cheaper MiMo-V2.6-Flash, it has released the model weights on Hugging Face and ModelScope. Releasing weights (open weights) means distributing the data of a trained model so that anyone can download it and run it in their own environment. The license is MIT, and commercial use is allowed.

What is Xiaomi MiMo?

Xiaomi MiMo is the name of the series of large language models (AI that understands and generates text) that Xiaomi develops in-house, and also the name of its development team. In Japan, Xiaomi is known as a smartphone maker, but it is also putting effort into developing AI models, and in April 2026 it released the previous-generation MiMo-V2.5 and MiMo-V2.5-Pro.

MiMo models can be used through an API for developers (a mechanism for calling them from other apps), as well as through Xiaomi MiMo Studio, where you can try them in a chat interface, Xiaomi MiMo Desktop for PCs and Xiaomi MiMo Code for programming assistance. Alongside this release, Xiaomi has announced that API requests to MiMo-V2.5 and MiMo-V2.5-Pro will be automatically switched to the V2.6 series from October 14, and the old model IDs will stop working on October 21.

Differences between Pro and Flash

Both MiMo-V2.6 models use an architecture called MoE (Mixture of Experts). Although the model as a whole is very large, each pass runs only some of its "expert" parts, so it can run fast and cheaply for its size. According to the model descriptions on Hugging Face, their scale is as follows.

MiMo-V2.6-Pro: 1.02 trillion parameters in total, of which 42 billion are used per pass
MiMo-V2.6-Flash: 309 billion parameters in total, of which 15 billion are used per pass

Both are "omnimodal" models that can understand four types of input in one model, text, images, video and audio, and their output is text. The amount they can read at once (context) is 1 million tokens, and a single output can be up to 128,000 tokens. Xiaomi says they were designed to handle entire sets of long program code and long work logs of AI agents (systems in which AI uses tools on its own to carry out tasks).

Xiaomi positions Pro as its top model for complex projects, long-running tasks, cybersecurity and research. Flash is a model that emphasizes the balance of cost and performance, aimed at tasks called repeatedly and high-volume processing. In addition, it has also provided MiMo-V2.6-Pro-UltraSpeed, an API-only model that outputs up to 20 times faster than Pro. For research, it has also released MiMo-V2.6-Distill-Qwen-9B, a small 9-billion-parameter model created by further training Alibaba's Qwen3.5-9B on data produced by MiMo.

Performance evaluation

On the overall index (Intelligence Index) of Artificial Analysis, a third-party evaluation site that compares AI models, MiMo-V2.6-Pro scored 46, ranking first among models with published weights. On its own site, Xiaomi gives the score to decimal places as 46.32 and calls it "the world's strongest open-source model at present." The Batch, a US newsletter covering AI topics, reports that this is up 20 points from the 26 points of the previous-generation MiMo-V2.5-Pro.

In Xiaomi's own published evaluations, the models showed scores close to Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 Sol on many tests, including automated programming tasks, operating computer screens and working in a terminal (a screen for operating a computer through text). For example, on Terminal Bench 2.1, which measures terminal tasks, Pro scored 89.9 and Flash 87.6. On the other hand, on the newer Terminal Bench 4.0, Pro scored 34.9, a large gap from Claude Opus 5's 49.0, and gaps with these top models also remain on some programming tests and on tests covering cyberattack techniques. All of these figures were measured by Xiaomi itself.

In addition, according to The Batch, Xiaomi has also disclosed the cost of the finishing stages such as reinforcement learning: about $2.6 million for Pro and about $900,000 for Flash.

What was released and pricing

On Hugging Face, the model weights, including the "RL" versions that have completed reinforcement learning (a method of learning from the results of trial and error), and a technical report have been released. Xiaomi says it has released not only the weights but also the reinforcement learning environments and the training code. The MIT license is a permissive license, and the models can be used commercially for free.

However, running them locally is not easy. Pro's launch example assumes a configuration of two GPU servers with SGLang, which is recommended, and eight GPUs in the vLLM example, so it is not something that can run on a personal computer. For most people, using it through Xiaomi's API or relay services such as OpenRouter is the realistic option.

Xiaomi's API pricing per million tokens is as follows (as of October 12, 2026).

Pro: $0.435 for input ($0.0036 with cache hits), $0.87 for output
Flash: $0.14 for input ($0.0028 with cache hits), $0.28 for output
Pro-UltraSpeed: $4.35 for input ($0.036 with cache hits), $8.7 for output

The API supports both the OpenAI and Anthropic calling formats. A flat-rate Token Plan is also available, and Xiaomi says it can be used directly from development tools such as Claude Code and Cline. There is no particular guidance on Japanese language support or availability for Japan.

CoAI's take

Until now, the representative high-performance open-weight models from China have been ones like DeepSeek, Qwen and Kimi. The big news this time is that Xiaomi, a smartphone maker, has put out a model that ranks first in an external evaluation. It can understand not only text but also images, video and audio together, and it can be used commercially under the MIT license, making it a strong option for companies that want to run AI in their own environments.

On the other hand, many of the evaluation figures were measured by Xiaomi itself. Running Pro on your own requires a large-scale GPU environment, and its usability in Japanese has not yet been sufficiently verified. A good first step is to try it through the API or MiMo Studio with material from your own work.

Sources

Back to news

Xiaomi Releases the Open-Weight AI Model MiMo-V2.6, Which Also Understands Images, Video and Audio, With Pro Topping Open Models in an External Evaluation