MiMo

MiMo-V2.5

1. Introduction

MiMo-V2.5 is a native omnimodal model with strong agentic capabilities, supporting text, image, video, and audio understanding within a unified architecture. Built upon the MiMo-V2-Flash backbone and extended with dedicated vision and audio encoders, it delivers robust performance across multimodal perception, long-context reasoning, and agentic workflows. Key features include:

• Hybrid Attention Architecture: Inherits the hybrid design from MiMo-V2-Flash, interleaving Sliding Window Attention (SWA) and Global Attention (GA) with a 5:1 ratio and 128 sliding window. This reduces KV-cache storage by nearly 6× while maintaining long-context performance via learnable attention sink bias.

• Native Omnimodal Encoders: Equipped with a 729M-param Vision Transformer (ViT) featuring hybrid window attention and a dedicated audio encoder initialized from the weights of MiMo-Audio, enabling high-quality image, video, and audio understanding.

• Multi-Token Prediction (MTP): Three lightweight MTP modules with dense FFNs accelerate inference via speculative decoding and improve RL training efficiency.

• Efficient Pre-Training: Trained on a total of ~48T tokens using FP8 mixed precision. The context window supports up to 1M tokens.

• Agentic Capabilities: Post-training incorporates SFT, large-scale agentic RL, and Multi-Teacher On-Policy Distillation (MOPD), achieving strong performance on agentic tasks and multimodal understanding benchmarks.

Model Summary

• Architecture: Sparse MoE (Mixture of Experts), 310B total / 15B activated parameters
• Context Length: Up to 1M tokens
• Modalities: Text, Image, Video, Audio
• Vision Encoder: 729M-param ViT (28 layers: 24 SWA + 4 Full)
• Audio Encoder: 261M-param Audio Transformer (24 layers: 12 SWA + 12 Full)
• Multi-Token Prediction (MTP): 329M parameters, 3 layers

2. Downloads

3. Evaluation Results

Multimodal Benchmarks

Coding & Agent Benchmarks

Long Context Benchmarks

4. Model Architecture

LLM Backbone

MiMo-V2.5's core language backbone inherits from the MiMo-V2-Flash architecture, a sparse MoE model with hybrid sliding window attention.

Vision Encoder

We train a dedicated MiMo ViT that adopts sliding-window attention to enable efficient visual encoding.

Window pattern notation: -1 = full attention, 0 = 1-D row window, 1 = 1-D column window.

Audio Encoder

Our audio encoder is initialized from the weights of MiMo-Audio-Tokenizer and further finetuned to support high-quality audio understanding.

5. Training Process

MiMo-V2.5 is trained on a total of ~48T tokens.

1. Text Pre-training: We collect diverse text data for pre-training the LLM backbone.
2. Projector Warmup: Short-duration warmup of multimodal projectors (audio and visual MLP projectors).
3. Multimodal Pre-training: High-quality multimodal data collected for large-scale pretraining.
4. SFT & Agentic Post Training: Supervised fine-tuning with diverse agentic data. During this stage, the context window is progressively extended from 32K → 256K → 1M.
5. RL & MOPD Training: Reinforcement learning for improving perception, reasoning, and agentic capabilities.

6. Deployment

Since inference engines are continuously being updated and optimized, this guide only provides deployment examples for reference. For the best performance, we strongly recommend following our referenced approach to get the latest best practices and optimal performance.

SGLang Deployment

For the best performance, we st