Xiaomi releases MiMo-V2.6-Pro and MiMo-V2.6-Flash open-source AI models with scaled RL training

Xiaomi has released and open-sourced the MiMo-V2.6 series. The company says the release focuses on scaling reinforcement learning (RL) compute on verifiable and complex tasks through exploration and feedback.

The series includes MiMo-V2.6-Pro and MiMo-V2.6-Flash, while Xiaomi is also rolling out MiMo-V2.6-Pro-UltraSpeed.

MiMo-V2.6-Pro and MiMo-V2.6-Flash

MiMo-V2.6-Pro and MiMo-V2.6-Flash are natively omnimodal models. MiMo-V2.6-Pro scored 46.32 on the Artificial Analysis Intelligence Index, surpassing Kimi K3 and Qwen3.8 Max. Xiaomi says the series retains the API pricing of the V2.5 series.

Xiaomi says it worked through the research and engineering challenges involved in RL training and streamed the production run live. In under six days, Flash and Pro each completed 30 RL steps across roughly 750,000 trajectories.

The results included:

• Training cost: About $0.85 million for Flash and $2.62 million for Pro
• Average pass rate: Increased by 25% and 12% in relative terms, respectively
• DeepSWE v1.1: Scores increased from 48.8 to 65.68 for Flash and from 58.4 to 72.57 for Pro

DeepSWE v1.1 is a held-out long-horizon software engineering benchmark. Xiaomi says RL continued improving throughout the run and generalized beyond the training distribution.

The company scaled RL compute across three areas:

• Larger batches: A fully asynchronous architecture uses 1,568 samples per update, supports up to 1 million-token context and processes 3.5–3.7 billion tokens per step.
• More tasks: The training suite covers coding, general agents, visual and cyber tasks across several harnesses.
• More grader compute: The approach provides more precise and diverse reward signals for long-horizon RL tasks, helping reduce paths and tokens per task.

As the run scaled, Xiaomi froze the router to suppress training drift. Its defenses against reward hacking include reward design, adversarial evaluation, anomaly detection and cross-checking between verifiers, with the company saying these measures improve training stability and reward reliability.

For agentic RL across different tasks, Xiaomi developed a unified trajectory representation and penalty mechanism, high-concurrency interaction across agent frameworks, a decoupled control and data plane, stabilized per-task sampling ratios and optimized training and inference engines.

Xiaomi is also open-sourcing the technical report, training environments and RL code for reproduction and verification.

From coding to interactive worlds

Beyond its RL training work, Xiaomi also highlights how MiMo-V2.6 applies these capabilities to multimodal and interactive tasks. The model adds 3D spatial reasoning, multimodal perception and computer-use capabilities (CUA), which Xiaomi refers to as “Vibe World” when describing the expansion from natural-language programming to interactive worlds.

For game development, the model can take an image, video or text prompt, break it into tasks and coordinate multiple agents to:

• Build 3D scenes and interaction logic
• Perform visual verification
• Refine results using rendered outputs
• Produce a runnable interactive world

It can also generate 3D objects and scenes in Blender from text descriptions or reference images. Xiaomi says these outputs can be used as assets for animation, 3D printing and game development.

Visual design and video

MiMo-V2.6 can turn an instruction into a frontend interface or slide deck with structured layouts, components, interactive elements and animation. Xiaomi says it can also work with Figma and image and video generation tools while maintaining consistency in typography, colors and text and image placement.

Xiaomi presented three presentation examples created from one-line briefs:

• Internal engineering training on how an LLM learns
• A research note on the critical minerals supercycle
• A design review proposing three directions for a company’s first annual report

For video, the model can handle visual design, shot and motion sequencing, music composition and beat-synchronized editing. For educational content, Xiaomi showed it explaining concepts such as Fourier decomposition through animations, with MiMo-V2.5-TTS generating narration aligned with the visuals.

Music composition

The model’s multimodal capabilities also extend to music composition. Xiaomi presented examples covering orchestral and piano compositions to show how the model works with musical structure, instrumentation and arrangement.

In one example, MiMo-V2.6-Pro was asked to compose an orchestral piece using about 10 instruments. It wrote the score and converted it to MIDI, while another example involved two piano pieces in A minor, with one using a slow tempo and the other a moderate tempo.

Xiaomi says these examples show how the model handles instrument roles and relationships in orchestration as well as dynamics, texture and other composition elements.

Research applications

Xiaomi presented MiMo-V2.6 applications in materials design and mathematical formalization without RL specifically tailored to scientific research.

For materials research, Xiaomi’s experts tasked MiMo-V2.6-Pro with designing a new metal-organic framework (MOF) material capable of adsorbing PFASs. Over several rounds of prompting, the model worked through the following steps:

• Searched the web and reviewed relevant literature and patents
• Proposed hypotheses and assessed novelty
• Conducted “dry experiments” using open-source computational tools
• Automatically set up the simulation environment
• Calculated binding strengths between MOF materials and PFASs
• Identified candidates for the “wet-lab” stage

For mathematics, MiMo-V2.6-Pro assisted researchers with a Lean 4 formalization of the main theorem in Li and Yorke’s “Period Three Implies Chaos.” The theorem states that a continuous self-map of an interval with a period-three orbit has periodic orbits of every positive integer period and an uncountable scrambled set.

Using an exploration strategy developed by researchers and subagent collaboration, the model advanced the theorem’s formalization and proof. The resulting work included:

• More than 6,000 lines of Lean source code
• A complete proof verified by the Lean kernel
• No unfinished proof placeholders

Xiaomi says the model had received no Lean-specific post-training for this work.

Availability and pricing

MiMo-V2.6-Pro and MiMo-V2.6-Flash are available today through:

• Xiaomi AI Studio
• MiMo Code
• MiMo Desktop
• MiMo API Platform
• OpenRouter

MiMo Desktop has exited early access and is now available as its first official release with Pro and Flash built in. It also introduces MiMo-V2.6-Pro-UltraSpeed, offering up to 20x faster output than Pro at the same quality.

The early-access program continues for one more week, while existing users can switch to the new model names. On the API side, MiMo-V2.6 is live on the Xiaomi MiMo Open Platform with pricing unchanged from V2.5.

Pro can be called in UltraSpeed mode with up to 20x faster output, while the Token Plan is available for predictable, high-volume usage. The API pricing is: