Xiaomi releases MiMo-V2.6 open-weight models with agent focus

Xiaomi has released two open-weight, multimodal AI models that it says bring advanced agent capabilities to developers at far lower prices than many proprietary alternatives. The new MiMo-V2.6-Pro and MiMo-V2.6-Flash models can process text, images, audio, and video, support a context window of up to one million tokens, and are designed for work that requires many steps, such as coding, research, design, and computer use.

According to Xiaomi, MiMo-V2.6-Pro is its most capable model so far. The company says it scored 46.32 on Artificial Analysis’ Intelligence Index, placing it ahead of other open-weight models in the benchmark. Carl Franzen reports for VentureBeat that the model ties xAI’s Grok 4.7 in the index and is listed as the benchmark’s highest-scoring open-weight model.

The company is also introducing MiMo-V2.6-Flash, a smaller version intended for high-volume use. Xiaomi positions Flash as a lower-cost option that remains close to Pro on several of its internal and vendor-run agent benchmarks. Both models are released under the MIT license, allowing organizations and developers to download, adapt, and run the weights themselves.

Long-running tasks are the focus

Xiaomi trained the models with reinforcement learning, a method that rewards a model for completing tasks successfully. The company says Pro and Flash each completed 30 large reinforcement-learning updates across roughly 750,000 task trajectories in less than six days. Xiaomi puts the reported cost at about $2.62 million for Pro and $850,000 for Flash.

Rather than training models only on short prompts and answers, Xiaomi says it trained them on lengthy workflows. These can include planning a task, using software tools, testing results, correcting mistakes, and verifying the final output. The company says it used environments covering coding, general agent tasks, visual work, and cybersecurity.

A key challenge in such training is reward hacking, where a system finds a way to meet a test without truly completing the intended work. Xiaomi says it blocked access to potential answer sources, removed software artifacts from training environments, and used separate systems to search for loopholes. It also evaluated different successful solutions against one another, aiming to favor cleaner and more efficient work over patches that simply pass a test.

The company reports substantial benchmark gains after reinforcement learning. On DeepSWE v1.1, a benchmark for long software-engineering tasks, Xiaomi says MiMo-V2.6-Pro reached 71.9, up from 19.0 for MiMo-V2.5-Pro. These figures are company-reported, and Xiaomi does not claim to lead every evaluation against closed models such as Anthropic’s Claude Opus 5 or OpenAI’s GPT-5.6 Sol.

Lower API prices and broader creative ambitions

Xiaomi lists API pricing for Pro at $0.435 per million uncached input tokens and $0.87 per million output tokens. Flash costs $0.14 for input and $0.28 for output. The company also offers a Pro UltraSpeed mode, which it says can generate responses at up to 20 times the normal output speed.

Beyond coding, Xiaomi is promoting the models for creating websites, presentations, video, 3D scenes, and music. Its demonstrations include generating Blender assets, coordinating agents to build interactive 3D environments, and controlling a simulated robotic arm from camera input. Xiaomi also describes research experiments involving materials discovery and formalizing a mathematics theorem in Lean 4.

Those examples show the range Xiaomi is targeting, but they do not independently establish that the models can reliably complete such work without expert oversight. Still, the release is notable because Xiaomi is publishing more than model weights. The company says it is also open-sourcing training environments, reinforcement-learning code, and technical documentation, potentially giving other teams a closer look at how large-scale agent training is conducted.

Sources

• Xiaomi debuts open-weight omnimodal models MiMo-V2.6 Pro and Flash – Xiaomi MiMo
• MiMo-V2.6-Pro ties Grok 4.7 and beats GLM-5.3 on Artificial Analysis’ Intelligence Index – VentureBeat