Xiaomi MiMO-V2-Flash

Xiaomi MiMO-V2-Flash: Overview

Xiaomi MiMO-V2-Flash is a cutting-edge open-source Mixture-of-Experts (MoE) large language model (LLM) developed by Xiaomi, released on December 16, 2025. It represents a significant advancement in efficient AI inference, combining frontier-level performance with high-speed processing and low costs. Designed primarily for reasoning, coding, agentic tasks, and long-context applications, it builds on Xiaomi's earlier MiMO efforts and is positioned as a competitive alternative to models like DeepSeek-V3.2 and Claude Sonnet 4.5. The model is fully open-sourced under the MIT license, making it accessible for research and deployment.

Key Features and Technical Specifications

• Parameters: 309 billion total parameters, with only 15 billion active per forward pass—enabling massive scale without proportional compute demands.
• Context Window: Up to 256,000 tokens, supporting extended agent interactions, tool calls, and multi-turn conversations.
• Inference Speed: Up to 150 tokens per second, with a 2.0–2.6× speedup via native Multi-Token Prediction (MTP).
• Cost Efficiency: $0.1 per million input tokens and $0.3 per million output tokens (assuming a 3:1 input/output ratio).
• Training Data: Trained on 27 trillion tokens in FP8 mixed precision for efficiency.
• Modes: Supports a "hybrid thinking mode" for step-by-step reasoning or instant responses, toggleable by users.
• Multilingual Focus: Strong performance in Chinese and English, with emphasis on agentic workflows like web development and code generation.

Architecture

MiMO-V2-Flash employs a hybrid attention mechanism to balance efficiency and long-context capability:

• Sliding Window Attention (SWA) with an aggressive 128-token window for local focus.
• Global Attention (GA) interleaved in a 5:1 ratio (GA:SWA) across 8 hybrid blocks.
• Learnable Attention Sink Bias: Maintains performance over ultra-long contexts.
• Multi-Token Prediction (MTP) Module: A lightweight 0.33B-parameter addition per block using dense Feed-Forward Networks (FFNs) and SWA. It enables self-speculative decoding, accepting 2.8–3.6 tokens per draft for faster generation.
• Post-Training: Uses MOPD (Multi-Teacher Online Policy Distillation), an efficient on-policy method that distills knowledge from multiple expert teachers with <1/50th the compute of traditional Supervised Fine-Tuning (SFT) + Reinforcement Learning (RL) pipelines. This supports continuous self-improvement.

The base model (MiMo-V2-Flash-Base) is also released, allowing fine-tuning for specialized tasks.

Benchmarks and Performance Comparisons

MiMO-V2-Flash excels in open-source rankings, particularly in coding and agentic benchmarks, often matching or surpassing closed-source leaders like GPT-5 and Gemini 3.0 Pro. It ranks #1 among open-source models on SWE-Bench and leads in math/reasoning tasks like AIME 2025.

Here's a summary of key benchmarks (post-training version):

• Strengths: Tops open-source in SWE-Bench (coding resolution) and AIME (math), with long-context outperforming full-attention models like K2-Thinking.
• Base Model Benchmarks: MMLU-Pro (73.2), MATH (71.0), HumanEval+ (70.7), and 96.7% on NIAH-Multi (256K context).

Availability and Usage

• Download: Model weights on Hugging Face at XiaomiMiMo/MiMo-V2-Flash. Includes base model and 3-layer MTP weights.
• API Access: Free limited-time trial on Xiaomi's platform (platform.xiaomimimo.com) and AI Studio (aistudio.xiaomimimo.com).
• Inference: Optimized for SGLang framework (GitHub: sgl-project/sglang). Example launch: --tp-size 8 --enable-mtp --speculative-algorithm EAGLE. Use FP8 for efficiency.
• API Example (Chat Completion):

  {
  "messages": [{"role": "user", "content": "Your query here"}],
  "max_tokens": 1024,
  "temperature": 0.8,
  "top_p": 0.95
  }

  • For agentic/math tasks: Set temperature=0.3.
  • System Prompt: "You are MiMo, an AI assistant developed by Xiaomi. The current date is [date]."
  • Tool-Use: Persist reasoning history in multi-turn calls.
• Technical Report: Available on GitHub (XiaomiMiMo/MiMo-V2-Flash).

Unique Aspects and Applications

• Web Development: Generates fully functional HTML pages from prompts, e.g., simulating macOS interfaces with interactive Terminal and Finder.
• Agentic Excellence: Strong in multi-step workflows, tool calling, and browsing (e.g., 58.3% on BrowseComp with context management).
• General Assistance: Handles philosophical discussions, creative writing, and everyday tasks with structured reasoning.
• Limitations: May underperform on very long contexts without bias tweaks; requires significant GPU resources (e.g., 8x TP for inference); agentic tasks need history persistence.

MiMO-V2-Flash democratizes high-performance AI through its open-source nature and efficiency, making it ideal for developers building agents, coders, and researchers in multilingual/long-context scenarios. For hands-on trials, start with the Hugging Face space or API. If you need code examples or comparisons to specific models, let me know!