Xiaomi's MiMo team recently released and open-sourced the MiMo-V2.6 series of models. The series includes two native full-modality models, MiMo-V2.6-Pro and MiMo-V2.6-Flash, with a focus on exploring how scaling up reinforcement learning can let models continuously improve through ongoing trial, error, and feedback.

Xiaomi says MiMo-V2.6 is a significant step in its exploration of Recursive Self-Improvement (RSI). Unlike simply expanding the volume of pretraining data, MiMo-V2.6 channels more computing power into verifiable, complex task environments, allowing the model to improve by completing tasks, receiving feedback, and correcting itself.

Large-Scale Reinforcement Learning Training Completed in Six Days

MiMo-V2.6's reinforcement learning training ran for six days. Xiaomi says these six days were a concentrated showcase of the team's long-term foundational research and engineering practice, backed by roughly half a year of work building the training system and running experiments.

During training, MiMo-V2.6-Pro and MiMo-V2.6-Flash each completed 30 rounds, generating around 750,000 trajectories in total. The training costs for the two models were about $2.62 million and $850,000 respectively.

After large-scale, multi-task reinforcement learning, MiMo-V2.6-Pro reached a level close to Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks, while MiMo-V2.6-Flash surpassed the previous-generation MiMo-V2.5-Pro overall.

On the training tasks, the average pass rates for Pro and Flash rose by about 12% and 25% respectively. On DeepSWE v1.1, a long-horizon software engineering benchmark not used in training, Pro's score climbed from 58.4 to 72.6 and Flash's from 48.8 to 65.7.

Xiaomi argues these results show that reinforcement learning not only improves a model's performance on training tasks but also boosts sample efficiency, sustained improvement, and generalization to new tasks.

Pro Model Reaches a Composite Intelligence Index Score of 46

MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index. Xiaomi says this beats Kimi K3 and Qwen3.8 Max, making it one of the stronger open-weight models by overall performance today.

That said, Xiaomi also notes that MiMo-V2.6-Pro still trails top closed-source models such as Claude Fable 5.1 and GPT-6 Astra.

On pricing, the MiMo-V2.6 series keeps the same API rates as the MiMo-V2.5 series. Xiaomi says that with improved capability and unchanged prices, MiMo-V2.6 further improves the balance between intelligence and usage cost.

By the company's own account, at a comparable level of intelligence, MiMo-V2.6-Pro costs about one-twentieth to one-sixtieth as much as similar overseas models.

Scaling Reinforcement Learning Along Three Directions

MiMo-V2.6's reinforcement learning training scaled up mainly along three directions.

The first is increasing batch size and training throughput. Through large-batch sampling and a fully asynchronous training architecture, each update uses 1,568 samples and supports contexts of up to 1 million tokens. A single training step processes roughly 3.5 billion to 3.7 billion tokens.

The second is adding more task types and greater environment complexity. The training system covers domains including code, general tasks, visual tasks, and cybersecurity, and introduces multiple harnesses so that different capability dimensions can improve together within one training framework.

The third is adding more computing power for the scorers. Through relative comparison within groups, the system provides more accurate and diverse reward signals for long-horizon reinforcement learning tasks, guiding the model to complete tasks along shorter paths and with fewer tokens.

As training scaled up, the model also became more prone to problems such as expert load drift and reward hacking. To improve training stability, the team froze the router of the MoE model and built an anti-reward-hacking mechanism spanning reward design, adversarial evaluation, anomaly detection, and cross-verification of verifiers.

Optimizing the Training System for Multi-Agent Tasks

For multi-agent reinforcement learning in large-scale mixed tasks, the MiMo team designed a unified trajectory representation and penalty mechanism to further refine the learning signal.

This mechanism supports interaction among multiple agent frameworks in high-concurrency environments and decouples the control plane from the data plane to make it easier to migrate and handle large-scale training trajectories. The system also controls the sampling ratio of different tasks within mixed batches, improving the efficiency and consistency of the training and inference engines.

Xiaomi says the related training framework and supporting resources have been open-sourced, so researchers can build on the publicly available models, task environments, and training code to continue exploring reinforcement learning algorithms, reward mechanisms, and the design of agent harnesses.

From Vibe Coding to Vibe World

MiMo-V2.6 further integrates 3D spatial reasoning, multimodal perception, and computer operation, extending natural-language-driven programming from web pages and code generation to interactive world building.

In 3D open-world game development, users can input images, video, or text requirements, and the model breaks the task into multiple subtasks, then uses multi-agent collaboration to build the 3D scene, program the interaction logic, and carry out visual verification.

The model can also build objects and scenes in Blender based on text descriptions or reference images, generating 3D assets ready for animation, 3D printing, and game development.

In embodied AI simulation environments, MiMo-V2.6 can take in multi-view camera feeds directly, reason and make decisions continuously, and use a visual feedback loop to control a Franka Panda robotic arm to complete tasks such as grasping objects, matching colors, and precise placement.

Further Strengthening Computer Use

MiMo-V2.6 also strengthens its Computer Use capability, combining multimodal perception with natively trained operational skills.

The model can understand complex graphical interfaces, call office and productivity tools to perform information retrieval, content editing, and data processing, and use visual feedback to check task results, pinpoint problems, and adjust subsequent actions.

This means the model does more than generate text or code; it can execute continuous tasks within real software interfaces, forming an "observe–act–check–correct" loop.

Contributing to Materials Science and Mathematical Formalization Research

Xiaomi says that even without reinforcement learning specifically tailored to scientific research tasks, MiMo-V2.6 has already shown potential across several research fields.

In materials research, MiMo-V2.6-Pro helped researchers design metal-organic framework materials for adsorbing per- and polyfluoroalkyl substances (PFAS).

Across multiple rounds of interaction, the model was able to retrieve and organize relevant papers and patents, propose research hypotheses, assess the novelty of design approaches, and call open-source computational tools to set up simulation environments, calculate the binding strength between candidate materials and PFAS, and select materials more worth validating experimentally.

In mathematical formalization, MiMo-V2.6-Pro helped researchers use Lean 4 to complete a full formalization of the "period three implies chaos" theorem from the classic paper by Li and Yorke.

The project ultimately produced more than 6,000 lines of Lean source code, verified by the Lean kernel with no unproven placeholders left behind. Xiaomi says the case did not use a model specially trained for Lean tasks, demonstrating MiMo-V2.6's potential in complex formal-proof tasks.

Supporting Creation Across Code, Design, Video, and Music

The MiMo-V2.6 series also strengthens code-driven content creation, spanning front-end web pages, Figma design files, presentations, SVG, video, and music.

In front-end and slide-deck creation, the model can generate complete pages and presentations from simple instructions, including structured layouts, interactive components, animation effects, and fonts, color schemes, and text-image composition.

In video creation, the model can handle visual design, shot arrangement, motion effects, soundtrack synthesis, and rhythm matching according to user needs. For explainer videos, it can also turn abstract concepts such as Fourier decomposition and convex hulls into accessible explanations and coherent animations, and call MiMo-V2.5-TTS to generate narration synced to the visuals.

In music creation, MiMo-V2.6-Pro can compose orchestral works featuring multiple instruments and automatically convert the generated score into MIDI format. Xiaomi says this reflects the model's grasp of instrument roles, arrangement relationships, melody creation, and overall composition.

Desktop Client and API Service Launched Alongside

With the release of MiMo-V2.6, Xiaomi is also launching the MiMo Desktop client and a membership service.

MiMo-V2.6-Pro also offers an UltraSpeed high-speed mode, which the company says can reach up to 20x inference speed, aimed mainly at use cases sensitive to real-time interaction and response latency.

The MiMo-V2.6 series is now available on Xiaomi's MiMo open platform, with API prices unchanged. Users can call the following models through the platform:

  • mimo-v2.6-pro
  • mimo-v2.6-flash
  • mimo-v2.6-pro-ultraspeed

Xiaomi says the existing invite-only testing will end in one week, and users who already have testing access can continue using the service after switching the model name.

Fully Open-Sourcing Model Weights and Reinforcement Learning Resources

In this release, Xiaomi not only opened up the model weights and technical report for MiMo-V2.6-Pro and Flash, but also released MiMo-V2.6-Distill-Qwen-9B along with supporting reinforcement learning resources.

The open-sourced content includes:

  • more than 7,000 high-quality reinforcement learning task environments;
  • task collections targeting software engineering, vulnerability reproduction, knowledge-intensive work, and web design and development;
  • an end-to-end reinforcement learning framework built on verl, uni-agent, and mini-swe-agent;
  • a complete training pipeline covering environment interaction, trajectory collection, reward evaluation, and policy optimization;
  • lightweight harnesses for building different system prompts, tools, and context management schemes.

In reinforcement learning experiments with Distill-Qwen-9B, the SWE-bench Verified score rose from 61.1 to 66.2, MiMo Cyber Bench from 31.3 to 47.0, Terminal Bench 2.1 from 37.1 to 52.8, and MiMo Visual Coding from 64.0 to 72.4.

Xiaomi says it hopes that by opening up the models, training code, and task environments, it can give researchers a unified experimental foundation and advance research into large-scale reinforcement learning, agent training, and model self-improvement.