MiMo-Pro

|  |  |  |  |

 |   |   | 

MiMo-V2.5-Pro

MiMo-V2.5-Pro is an open-source Mixture-of-Experts (MoE) language model with 1.02T total parameters and 42B active parameters. It utilizes the hybrid attention architecture and 3-layers Multi-Token Prediction (MTP) introduced in MiMo-V2-Flash, with up to 1M tokens context length.

1. Introduction

MiMo-V2.5-Pro is our most capable model to date, designed for the most demanding agentic, complex software engineering, and long-horizon tasks. It sustains complex trajectories spanning thousands of tool calls with strong instruction following and coherence over a 1M-token context window. Key features include:

• Hybrid Attention Architecture: Interleaves Sliding Window Attention (SWA) and Global Attention (GA) with a 6:1 ratio and 128 sliding window. This reduces KV-cache storage by nearly 7x while maintaining long-context performance via learnable attention sink bias.
• Multi-Token Prediction (MTP): Equipped with three lightweight MTP modules using dense FFNs. This triples output speed during inference and will be good to accelerate rollout in RL training.
• Efficient Pre-Training: Trained on 27T tokens using FP8 mixed precision and native 32k seq length. The context window supports up to 1M tokens.
• Agentic Capabilities: Post-training utilizes SFT, large-scale agentic RL and Multi-Teacher On-Policy Distillation (MOPD), achieving superior performance on the most demanding agentic, complex software engineering, and long-horizon tasks.

2. Model Downloads

3. Evaluation Results

Base Model Evaluation

Long-context Evaluation

GraphWalks is a long-context benchmark from OpenAI that fills the prompt with a directed graph of hex-hash nodes and asks the model to run a breadth-first search (nodes exactly at depth N) or list a node's parents. We evaluate across the full 32k–1M input-token span and apply the same evaluation fixes described by Anthropic.

MiMo V2.5 Pro delivers a major leap in long-context reasoning. Past 128k, V2 Pro degrades rapidly and collapses to 0.00 at 1M on both subtasks, while V2.5 Pro still scores 0.56 BFS / 0.92 Parents at 512k and 0.37 / 0.62 at 1M.

4. Model Architecture & Training Process

MiMo-V2.5-Pro addresses the quadratic complexity of long contexts by interleaving Local Sliding Window Attention (SWA) and Global Attention (GA). Unlike traditional speculative decoding, our MTP module is natively integrated for training and inference.

Model Summary

Training Process

For post-training, MiMo-V2.5-Pro adopts the three-stage post-training paradigm introduced in MiMo-V2-Flash to achieve exceptional performance. The paradigm begins with Supervised Fine-Tuning (SFT) to build strong, foundational instruction-following skills using curated data pairs. Next, in the Domain-Specialized Training stage, diverse teacher models — ranging from math and safety to complex agentic tool-use — are individually optimized using domain-specific RL rewards. Finally, the process culminates in Mult