Xiaomi MiMo-V2.6: The Open Model That Learns From Work

Picture a coding agent at the end of a long debugging session. It has a repository, a failing test, a screenshot from a bug report, and a terminal full of clues. The hard part is not producing one clever answer; it is keeping a plan alive across many steps, checking each result, and recovering when a tool goes wrong. Xiaomi’s MiMo-V2.6, released on September 22, 2026, is aimed at that messier kind of work.

An agent is an artificial intelligence model connected to tools such as a terminal, browser, or office application. A multimodal model can work with more than text—in MiMo-V2.6’s case, text, images, video, and audio in one model family. The release includes MiMo-V2.6-Pro for maximum capability, MiMo-V2.6-Flash for a speed-and-cost balance, and a Pro-UltraSpeed rollout for faster generation. The interesting story is not only that Xiaomi built a larger model. It is that the model was trained around long, checkable tasks.

The big bet: learn from attempts

Reinforcement learning, usually shortened to RL, trains a model through attempts and feedback. The model takes an action, receives a reward—a numerical signal indicating how well it did—and adjusts its future behavior. For software work, a verifier can run tests. For a computer-use task, a harness, meaning the surrounding software that supplies tools and checks the final state, can inspect whether the job was completed.

A useful teaching sketch looks like this:

for step in range(30):
 attempts = model.try_tasks(prompts, tools=harness)
 scores = verifier.check(attempts)
 model.learn_from(attempts, scores)

This is not Xiaomi’s training code. It shows the rhythm: produce many complete attempts, score them, and use the feedback in the next update. A trajectory is one full attempt, including the model’s reasoning, tool actions, corrections, and final result. One RL step is a large update cycle, not one user question.

In less than six days, MiMo-V2.6-Pro and Flash each completed 30 RL steps across roughly 750,000 trajectories. Xiaomi reports training costs of about $2.62 million for Pro and $850,000 for Flash. Average pass rates on the training tasks rose by 12% and 25% respectively, while the held-out DeepSWE v1.1 software-engineering score climbed from 58.4 to 72.6 for Pro and from 48.8 to 65.7 for Flash.

Held-out means the tasks were kept outside the training set. Improvement there is a stronger sign of generalization—the ability to carry a learned strategy into unfamiliar work—than improvement on exercises the model has already seen. It still is not proof that every real project will behave the same way, especially when a model can exploit a weak grader. That shortcut is called reward hacking.

Two models, two jobs

MiMo-V2.6-Pro is the flagship for difficult, long-horizon work, meaning tasks that require many dependent steps over time. Flash is designed to preserve much of that ability while reducing response time and cost. Xiaomi also advertises Pro-UltraSpeed, claiming up to 20 times faster output at the same quality. That figure should be treated as a vendor claim rather than a universal promise: hardware, load, prompt length, and streaming settings all affect latency, the wait before output arrives.

The model documentation adds an important architectural detail. Both variants use a Mixture of Experts, or MoE, architecture. Instead of sending every token through one giant dense network, a routing layer selects a small group of specialist blocks. Pro is described as having about 1.02 trillion total parameters and 42 billion active parameters; Flash has about 309 billion total and 15 billion active. A parameter is a learned numerical weight. Active parameters reduce the work needed for each token, but they do not make a trillion-parameter model fit inside an ordinary laptop.

Both models support a context window of up to one million tokens. A token is a small unit used to represent text or other input, while a context window is the amount of material the model can consider in one request. That much room is useful for large repositories, long tool traces, and multi-session tasks, although more context does not guarantee perfect attention to every detail.

One model, many senses

Native omnimodal means the model is trained to understand text, images, video, and audio as part of one system rather than handing each input to an unrelated specialist. That matters when the evidence is scattered: a bug may be described in text, shown in a screenshot, and reproduced in a screen recording.

Xiaomi’s examples push beyond image captioning. MiMo-V2.6 can coordinate agents to build a three-dimensional, or 3D, scene, write its interaction logic, inspect rendered results, and revise the work. It can also use multi-view camera frames in a simulated environment to guide a Franka Panda robotic arm through grasping and placement. This is a closed loop: the model acts, sees the result, and uses that feedback to choose the next action. The demonstration is promising, but a simulation is not the same as reliable autonomy around real people and machinery.

How to read the benchmark table

A benchmark is a standardized test suite, not a universal measure of intelligence. MiMo-V2.6’s reported results show why several tests are more informative than one headline number:

The scores should be compared within a row, not across rows. CyberGym and ExploitGym, for example, produce very different pictures: Pro scores 94.0 on CyberGym but 17.8 on ExploitGym, while Flash scores 95.1 and 6.0. That spread is a useful warning. Strong performance in one carefully designed environment does not certify broad security ability.

Xiaomi also reports a 46.32 score for Pro on the Artificial Analysis Intelligence Index and frames it as a new open-model cost-performance point. Because the family keeps the V2.5 application programming interface, or API, pricing, the claim is about the Pareto frontier—the set of choices where getting more intelligence normally costs more money. MiMo-V2.6’s pitch is that the curve moved outward: more capability for a similar API bill.

Open does not mean small

Open-sourcing the weights and training materials makes the release more useful to researchers, but it does not turn Pro into a desktop download. Serving a model of this scale requires serious multi-GPU infrastructure, where GPU means graphics processing unit, along with careful memory management and software that understands sparse experts and multimodal input. A smaller distilled checkpoint is more approachable; distillation means training a compact model to imitate the behavior of a larger teacher.

The broader idea is the lasting part of MiMo-V2.6. Xiaomi is treating environments, graders, tool use, and training logs as part of the model’s development—not as plumbing hidden behind a chat window. If that recipe continues to transfer from coding to visual work, cybersecurity, and everyday automation, open models may improve through better practice, not only through more parameters.

MiMo-V2.6 is best understood as a public experiment in learning from work. The benchmark scores are memorable, but the more important test is whether a model can attempt, inspect, recover, and try again without losing the thread. That is where a chatbot starts to look like an agent.