MiMo-V2.6-Pro
Xiaomi’s 1T parameter multimodal model with 1M token context
MiMo-V2.6-Pro is a large language model developed by Xiaomi. Released as part of the MiMo-V2.6 series, it uses reinforcement learning to improve its own capabilities across text, image, video and audio. The model features one trillion total parameters and supports very long contexts for complex agent tasks.
- 1.02 trillion total parameters with 42 billion active
- Supports text, image, video and audio inputs
- One million token context window for long tasks
- Uses groupwise agentic grading for self-improvement
- Trained with mixed reinforcement learning across domains
What it is
MiMo-V2.6-Pro is the leading checkpoint in the MiMo-V2.6 series created by Xiaomi. It is designed to scale reinforcement learning towards self-improvement by expanding compute and environment diversity. The model handles text, images, video and audio within a single system. It supports a context window of one million tokens, allowing it to process long code repositories and multi-session agent runs. With 1.02 trillion total parameters, it uses a sparse architecture to activate only 42 billion per step.
How it came about
The model builds on the MiMo-V2.5 series. Xiaomi developed this version to focus on scaling reinforcement learning compute alongside grader compute. The team introduced a method called You Only RL Once, which mixes coding, general agents, visual tasks and cybersecurity into a single training run. This approach allows strategies learned in one domain to transfer to others. The release includes both Pro and Flash variants, with weights available on platforms like Hugging Face and ModelScope.
Inside the model
The architecture uses a sparse Mixture of Experts design with 384 routed experts, activating eight per token. It has 70 transformer layers, mixing local sliding window attention and global attention. A separate vision encoder processes images and video, while an audio tokenizer handles speech. The model employs Groupwise Agentic Grading to create a self-improvement loop. This system compares multiple outputs to build task-specific rubrics and reward higher-quality solutions. Multi-Prefix Multi-Teacher On-Policy Distillation helps train decision points without regenerating entire histories.
What it can do
The model performs well in coding benchmarks such as DeepSWE and ProgramBench. It shows strong results in general agent tasks like AutomationBench and Toolathlon. In cybersecurity tests, it achieves high scores on CyberGym and MiMo Cyber Bench. Visual agent capabilities are measured via MiMo VisualCoding. The system aims to reduce token usage per task by steering towards shorter paths. It handles long-horizon tasks through its extended context window.
How to use it
Weights are available for download on Hugging Face and ModelScope. Users can deploy the model using SGLang or vLLM frameworks. Docker images are provided for easier setup. The model is also accessible via the Xiaomi MiMo Open Platform API, AI Studio and OpenRouter. Recommended sampling settings include a temperature of 1.0 and top_p of 0.95.
Safety and testing
Training includes environment hardening and adversarial screening to prevent reward hacking. Verifier cross-checks ensure the self-improvement loop remains honest. Usage restrictions prohibit unlawful, fraudulent or malicious purposes. Military applications and malicious cyber activities like malware creation are banned. Users must not exploit personal data or automate third-party platforms without authorization. High-risk autonomous actions require human oversight.
Questions people ask about MiMo-V2.6-Pro
What makes MiMo-V2.6-Pro different from previous versions?
It scales reinforcement learning compute and grader compute together. The model uses a mixed training run across coding, agents, visual and cybersecurity tasks. This allows capabilities to reinforce each other rather than being trained separately.
How many parameters does the model have?
The model has 1.02 trillion total parameters. However, it uses a sparse Mixture of Experts architecture that activates only 42 billion parameters for each token processed. This makes it more efficient during inference.
Can it process video and audio?
Yes, it is a native omnimodal model. It accepts text, images, video and audio as inputs. It uses a specific vision encoder for visual data and an audio tokenizer with a patch encoder for sound processing.
Is the model open source?
The weights are available for download on Hugging Face and ModelScope. It is not fully open source in terms of training data, but the model weights are accessible. Users must follow specific licence restrictions regarding lawful use.
Good reading and sources
There is no English Wikipedia article on MiMo-V2.6-Pro yet, so the facts on this page come from the sources listed above: huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.