Xiaomi MiMo-V2.6 - Xiaomi's Open-Source Omni-Modal Model Series

Technical Principles of Xiaomi MiMo-V2.6

• Large-Scale Agentic RL: The model continuously trial-and-errors across multi-task environments such as Code, General, Visual, and Cyber, using verifiable results as reward signals; it completes approximately 750,000 trajectories in under 6 days, relying on large batches, a fully asynchronous architecture, and 1M context training to push single-step updates to 2.7–3.7B tokens, thereby improving sample efficiency and out-of-sample generalization.
• Reward and Anti-Cheating Mechanisms: Provides finer rewards for long-horizon tasks through relative comparison within groups, and suppresses Reward Hacking through reward design, adversarial evaluation, anomaly detection, and cross-validation by verifiers; meanwhile, it freezes the MoE Router to suppress expert load drift and ensure stability in large-scale training.
• Unified Trajectory and System Decoupling: Uses unified trajectory representation and penalty mechanisms to refine learning signals, supporting high-concurrency interaction in multi-agent frameworks; decouples the control plane from the data plane to facilitate massive trajectory transfer, stabilizes task ratios in mixed batches, and optimizes consistency between training and inference engines.
• From Omni-Modal to World Operation: Combines language, vision, 3D spatial reasoning, and native action capabilities, enabling the model not only to generate code but also to correct behavior in a closed loop based on visual feedback in Blender, game engines, simulation environments, and GUIs.

Follow WeChat and reply "open source" to join theAI Open-Source Project Communication Group

How to Use Xiaomi MiMo-V2.6

Method 1: MiMo Desktop Client

• Open XiaoMi MiMo Desktop to download and install MiMo Desktop.
• After logging into your account, choose a membership subscription, or configure your own API Key in Settings.
• Switch between MiMo-V2.6-Pro / Flash in the model list; for real-time scenarios, you can enable UltraSpeed.
• Directly initiate programming, PPT/webpage/SVG/video/music creation, research, or computer operation tasks using natural language, and upload images, videos, files, and other materials.
• Through follow-up questions, screenshot feedback, requesting retries, or phased confirmation, let the model iterate and correct based on visual/log results.

Method 2: MiMo Open Platform API

• Go to the Xiaomi MiMo Open Platform, create an application, and copy the API Key.
• Select the model MiMo-V2.6-Pro / Flash, and enable UltraSpeed ultra-high-speed mode as needed.
• Pass in the model name, messages, tool/multimodal inputs, and return format according to the platform API documentation.
• For Agent tasks, integrate environment logs, test pass rates, screenshots, or verifier feedback to form an "execute—check—correct" loop.
• First conduct small-traffic canary validation of price, latency, success rate, and stability, then gradually replace it in the production pipeline.

Core advantages of Xiaomi MiMo-V2.6

• RL at scale: Uses verifiable complex tasks as the core for large-scale Agentic RL, about 750,000 trajectories in 6 days, with empirical evidence of out-of-sample generalization.
• Outstanding Agent capabilities: Benchmarks against top closed-source models on multiple tasks such as software engineering, terminal, web, vision, and Cyber.
• From full modality to action: Covers text, images, video, GUI, 3D, Blender, robotic arms, and Computer Use.
• Strong cost-effectiveness: The article states that at the same intelligence level, the price is about 1/20–1/60 of overseas models, and API prices remain unchanged.
• Engineering stability design: Uses MoE Router freezing, Reward Hacking defenses, and a unified trajectory mechanism to support stable training.
• Amphibious in creation and research: Can do frontend/PPT/video/music, and can also participate in materials screening and Lean 4 formal proofs.

Project address of Xiaomi MiMo-V2.6