MiMo-V2.6-Pro-RL

Generic Info

• Publisher: Xiaomi MiMo
• Release Date: September 2026
• Architecture: Omnimodal Mixture-of-Experts with Multi-Harness RL Scaling
• Context Window: 1,000,000 tokens (1M native)
• License: MIT License
• Modalities: Text, High-Resolution Vision, Video, Audio in a single forward pass
• Key Capabilities: Groupwise Agentic Grading, Long-Horizon Coding, Cybersecurity, Self-Correction

MiMo-V2.6-Pro-RL represents a major leap in open-weight reinforcement learning toward self-improvement. By scaling RL compute, environment diversity, and agentic grading together, MiMo achieves the #1 ranking across independent open-weight intelligence benchmarks. Its unified "You Only RL Once" training mixes coding, tool use, visual workflows, and cybersecurity into a cohesive agentic model capable of executing multi-session repository-wide runs.

Hello World Guide

Run MiMo-V2.6-Pro-RL locally using Hugging Face transformers or vLLM.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "XiaomiMiMo/MiMo-V2.6-Pro-RL"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

prompt = "Analyze this complex algorithmic puzzle, verify constraints, and synthesize a proof."
messages = [
    {"role": "system", "content": "You are MiMo-V2.6-Pro, an omnimodal RL-trained reasoning assistant."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.6,
    top_p=0.95
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

Industry Usage