MiMo-Pro-RL

|  |  |  |  |  |

 |   |   | 

MiMo-V2.6-Pro-RL

Scaling Reinforcement Learning Toward Self-Improvement

1. Introduction

MiMo-V2.6-Pro-RL is the flagship checkpoint of the MiMo-V2.6 series. The series is built to scale reinforcement learning toward self-improvement — scaling RL compute, environment diversity, and grader compute together, so the model keeps expanding its capability frontier through exploration and feedback. Key features include:

• Native Omnimodal + Long Horizon: Text, image, video, and audio in one model; 1M tokens for long repositories, tool traces, and multi-session agent runs.
• You Only RL Once: One mixed RL run across coding, general agents, visual, and cybersecurity — not separate per-domain runs. Tasks and multiple harnesses are mixed in the same batch so capabilities reinforce each other and strategies transfer to harnesses never seen in training.
• Scaling RL Compute: Fully asynchronous Group Relative Policy Optimization (GRPO) on very large batches — 1,568 prompts × 16 rollouts per step, billions of tokens per update.
• Groupwise Agentic Grading (Self-Improvement Loop): Binary pass/fail cannot rank passing solutions, so the reward signal itself is scaled. An agentic grader compares rollouts within each group: Groupwise Reward Synthesis (GRS) builds task-specific rubrics offline from contrasting rollouts and fuses rubric quality with test outcomes; Groupwise Advantage Redistribution (GAR) ranks passing trajectories online and moves advantage toward higher-quality solutions. Judged against the policy’s own samples, this closes a self-improvement loop and steers toward shorter paths and fewer tokens per task.
• Aligned RL: Cold start from self-correction — the model reflects on and rewrites its own misaligned turns into grounded next steps. Throughout RL, environment hardening, adversarial screening, and verifier cross-checks keep the loop honest against reward hacking.
• Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2): After mixed RL, MOPD2 combines autonomous student rollouts with prefix-conditioned single-turn rollouts (Teacher-Prefix and SFT-Prefix), reusing histories from teacher trajectories and SFT demonstrations so decision points train without regenerating preceding turns — extending capabilities to hard-to-verify tasks.

Model Summary

• Architecture: Sparse MoE (Mixture of Experts), 1.02T total / 42B activated parameters
• Context Length: 1M tokens
• Modalities: Text, Image, Video, Audio
• Vision Encoder: 681M-param MiMo ViT (28 layers: 24 SWA + 4 Full)
• Audio Encoder: 308M AudioTokenizer + 127M audio patch encoder
• Multi-Token Prediction (MTP): 5-layer speculative decoder

Figure 1. MiMo-V2.6 architecture.

2. Downloads

3. Evaluation Results

4. Model Architecture

LLM Backbone

The first Transformer block uses global attention with a dense FFN. Remaining blocks interleave local SWA and GA; both use sparse MoE FFNs without shared experts.

Vision Encoder (MiMo ViT)

Audio Encoders

AudioTokenizer encoder: 24 layers (12 SWA / 12 GA), hidden 1024, 20 RVQ codebooks, 308M parameters. Audio patch encoder: 6 layers, 127M parameters; four frames per patch (25 Hz → 6.25 Hz).

Speculative Decoder

5-layer SWA MTP drafter (DFlash-style). Predicts 7 subsequent tokens per forward pass for parallel verification.

5. Deployment

For best performance, follow the SGLang MiMo cookbook. Doc