Xiaomi’s MiMo-V2.6: The New Open-Source King of Multimodal Agents

To ensure the model scales effectively, Xiaomi introduced Groupwise Agentic Grading (GRS + GAR). This mechanism scales the reward signal into a self-improvement loop, allowing the model to grade its own agentic trajectories. Additionally, the Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2) technique was utilized to extend these capabilities into 'hard-to-verify' tasks where traditional reward functions often fail.

• YORO: Unified RL across coding, visual, and cyber domains
• GRS + GAR: Self-improving reward loops for agentic scaling
• MOPD2: Advanced distillation for complex, non-verifiable tasks

Performance: Outperforming the Giants

The benchmarks are nothing short of historic. MiMo-V2.6 Pro achieved a score of 46 on the Artificial Analysis Intelligence Index, making it the highest-rated open-weight model in existence. In head-to-head comparisons, the Pro version performs on par with industry leaders like Claude Opus 5 and GPT-5.6 Sol across the majority of agentic benchmarks.

In specialized domains, the numbers are even more impressive. On the DeepSWE v1.1 coding benchmark, it scored 71.9, while in the AutomationBench, it hit 53.1—surpassing Claude Opus 5's score of 50.3. For developers concerned with cybersecurity, the model's performance on CyberGym (94.0) and the MiMo Cyber Bench (80.2) demonstrates a level of technical proficiency that makes it a viable tool for security research and automated pentesting.

• DeepSWE v1.1: 71.9
• AutomationBench: 53.1 (vs Claude Opus 5: 50.3)
• OSWorld-Verified: 82.0
• Toolathlon-Verified: 76.9
• CyberGym: 94.0

API Pricing and Accessibility

Xiaomi is making it incredibly easy for developers to integrate MiMo-V2.6 into their stacks. While the model weights are open-source, the managed API provides a high-availability route for production-grade applications. The pricing structure is designed to be competitive with the most aggressive providers in the market.

The MiMo-V2.6 API features highly optimized caching. For high-frequency RAG applications, the 'Cache Hit' pricing is exceptionally low, allowing developers to maintain massive context windows without breaking the bank. For lightweight or latency-sensitive tasks, the MiMo-V2.6 Flash model offers an even more economical alternative.

• Open Weights: Full access to model weights and training code
• Optimized Caching: Significant cost reductions for repetitive context
• Developer Friendly: Technical reports and RL environments released

Use Cases: From Coding Agents to Cyber Defense

The versatility of MiMo-V2.6 makes it a Swiss Army knife for AI engineers. In the realm of software engineering, the MiMo Code harness can handle ultra-long, 200+ step tasks, outperforming existing tools like Claude Code. It can navigate complex repositories, implement logic, and verify its own output.

Beyond coding, its agentic capabilities allow for the creation of autonomous workflows. Imagine an agent that can watch a video of a UI interaction, write the corresponding automation script, test it in a virtual environment, and refine the code until it works perfectly. This level of cross-modal reasoning is the future of autonomous software and digital assistants.

• Agentic Coding: Long-horizon, multi-step software development
• Cybersecurity: Automated vulnerability detection and testing
• Multimodal RAG: Deep reasoning over video, audio, and text
• Autonomous Agents: Complex task execution across digital environments

Getting Started

Developers can jump into the MiMo ecosystem immediately. Xiaomi has released the model weights, the technical report, the RL environments, and the training code to the community. You can download the weights from official repositories to run locally on your own infrastructure or utilize the high-speed API for rapid prototyping.

For those looking to build specialized agents, the released RL environments provide a sandbox to fine-tune the model's behavior for specific domain-driven tasks, leveraging the same principles that made MiMo-V2.6 a benchmark leader.

• Access: Open weights available for local deployment
• Resources: Full technical report and training code released
• Sandbox: RL environments available for custom fine-tuning

API Pricing — Input: $0.435 per million tokens / Output: $0.87 per million tokens / Context: 1M tokens

Sources