Xiaomi open-sources MiMo-V2.6 together with the RL stack used to train it

Xiaomi has released its MiMo-V2.6 AI model series as open models you can run on your own servers. The series includes two large models, Pro and Flash, and a smaller 9B model derived from them. What stands out is that Xiaomi released more than the models: it also opened the stack it used to train them, including more than 7,000 training environments, end-to-end training code, ready-to-run Docker environments and its grading system. The company also shared what the final training phase cost. That level of detail is rare among large AI labs.

What MiMo-V2.6 is

MiMo-V2.6 is a model family built especially for long coding jobs, tool use, computer use and multi-step agent tasks. Besides text, the models understand images, video and audio, and they can work with up to 1 million tokens at once. That means they can hold a large code repository, long tool logs or an agent run spanning several sessions in view at the same time.

The series has three models:

• MiMo-V2.6-Pro: The flagship. It has more than 1 trillion parameters in total, but thanks to its mixture-of-experts (MoE) design, only a small share of them runs at each step.
• MiMo-V2.6-Flash: Roughly a third the size of Pro, faster and cheaper to run.
• MiMo-V2.6-Distill-Qwen-9B: Alibaba's Qwen3.5-9B model trained on data generated by MiMo. Xiaomi offers it as a starting point for anyone researching reinforcement learning.

How big it is and where it runs

Pro and Flash are too large for a home computer or a single graphics card. According to Xiaomi's deployment guide, Pro runs spread across two servers, each packed with data-center GPUs. Flash is smaller but still needs a multi-GPU server. Ready-made recipes are available for common serving software such as SGLang and vLLM.

For people who want to try it on their own hardware, the 9B model is the realistic option, since a model that size can run on a single powerful graphics card. Those who want to try the large models without setting them up can use Xiaomi's own API platform, AI Studio, the MiMo Desktop app or OpenRouter.

The real news: the training stack is open too

Companies that release open models usually share only the finished model. How it was trained, which tasks it practiced on and how success was measured often stay private. This time Xiaomi shared most of that:

• Training environments: More than 7,000 tasks across software engineering, vulnerability reproduction, knowledge work, web development and music composition. Results are checked with executable tests for coding, rule checks for security, rubrics for knowledge work and visual grading for web development.
• Training code: Xiaomi's adaptation of verl, ByteDance's open-source training library, plus its agent harness mimoagent and the uni-agent model gateway.
• Docker environments: Ready-made containers the tasks run in.
• Training records: Data showing how the training run progressed step by step.

This means a university lab or a company can apply Xiaomi's method to its own model, or try to reproduce the results itself.

How the model was trained

In reinforcement learning, a model attempts a task, gets a reward when it succeeds and updates itself based on that reward. According to Xiaomi's technical report, a few ideas stand out in how MiMo-V2.6 was trained:

• One mixed run: Coding, general agent work, visual tasks and cybersecurity were not trained separately. They were mixed in the same training run. According to Xiaomi, strategies learned in one area carry over to the others this way.
• Very large batches: Each training step used 1,568 tasks, each attempted 16 times. That adds up to billions of tokens per step.
• Ranking the passing answers too: Of two solutions that pass a task, one may be shorter and cleaner. A simple pass-or-fail score cannot see that difference. Xiaomi uses a grading agent that compares different attempts at the same task and rewards the better ones more. According to the company, this pushes the model toward shorter paths and fewer tokens.
• Guarding against cheating: Models sometimes find ways to fool the scoring system instead of solving the task. Xiaomi says it hardened its environments and cross-checked its graders against each other.

What the training cost

According to Xiaomi, this final training phase took less than six days. Each model was trained for 30 steps, with about 750,000 task attempts in total. The cost was about $850,000 for Flash and about $2.6 million for Pro. According to the technical report, roughly half of that went to the model attempting tasks, about half to the training itself and a small share to the grader.

These figures are not the total cost of the model. The pre-training phase, where the model was first trained on trillions of tokens, staff salaries and failed experiments are not included. Xiaomi also did not say which GPUs it used or for how many hours. Some translations circulating on X show these figures ten times lower, because the Chinese unit for ten thousand was lost in translation.

The results Xiaomi shared

According to the table on Xiaomi's model card, MiMo-V2.6-Pro comes close to closed frontier models such as Claude Opus 5 and GPT-5.6 Sol on many coding-agent, general-agent and computer-use benchmarks, and beats them on some. On exploit benchmarks, however, it trails them clearly. Flash comes close to Pro on many tests. The table also shows a big jump over the previous MiMo-V2.5-Pro. The 9B model does noticeably better than the Qwen3.5-9B it is based on, especially on agent tasks.

Who it is for

• Researchers: People working on reinforcement learning for agents get ready-made environments, training code and training records at this scale.
• Companies that want their own setup: Organizations that do not want their data leaving the building can run Flash or Pro on their own servers, if they have the hardware.
• Small teams: The 9B model offers a starting point for agent experiments on limited hardware.
• Developers: They can try the model in coding agents through the API or OpenRouter without setting it up.

The models are released under the MIT license, and the training environments and training code under Apache 2.0. Both allow commercial use. All the benchmark results and cost figures shared are Xiaomi's own, and some come from benchmarks the company built itself. Because the stack is open, others can now try to reproduce these results, so it is worth following evaluations from independent researchers as well.

Sources: Hugging Face: MiMo-V2.6-Pro-RL, Hugging Face: MiMo-V2.6-Flash-RL, Hugging Face: MiMo-V2.6-Distill-Qwen-9B, MiMo-V2.6 technical report, Hugging Face: training environments, GitHub: XiaomiMiMo/verl.