MiMo-V2.5 - Xiaomi's Omni-modal Agent Large Model Series

What is MiMo-V2.5

MiMo-V2.5 is a new-generation omni-modal Agent large model series launched by the Xiaomi Large Model Team, including four versions: V2.5, V2.5-Pro, V2.5-TTS, and V2.5-ASR. MiMo-V2.5 is a native omni-modal Agent that supports image, audio, and video understanding and action, with a context length of up to 1M tokens; MiMo-V2.5-Pro is designed for long-horizon complex tasks and can stably complete nearly a thousand rounds of tool calls. Its software engineering capabilities benchmark against Claude Opus 4.6 and GPT-5.4, with token efficiency improved by 42%~50%. The entire MiMo-V2.5 model series will soon be open-sourced.

Main Features of MiMo-V2.5

• MiMo-V2.5

  • Native Omni-modal Agent: Supports simultaneous seeing, listening, and reading, transforming understanding into action, covering multi-modal perception of images, audio, and video.
  • General Agent Scenarios: Capable of handling tasks from daily simple to medium complexity, with API costs reduced by about 50% compared to the previous generation.
  • Multi-modal Perception Upgrade: Cross-modal reasoning, video understanding, and chart analysis capabilities comprehensively surpass the previous generation MiMo-V2-Omni, approaching top closed-source models in evaluations such as VideoMME, CharXiv, and MMMU-Pro.

  MiMo-V2.5-Pro

  • Long-horizon Complex Tasks: Can stably complete professional-level work with nearly a thousand rounds of tool calls in a single session, with significantly enhanced instruction following and ambiguous instruction understanding.
  • Complex Software Engineering: Supports large codebase development, multi-language programming, code debugging, and performance optimization, and can independently deliver projects such as compilers and video editors.
  • Autonomous Project Delivery: Implemented a complete SysY compiler from scratch in Rust (4.3 hours, 672 calls, perfect score of 233/233 on the hidden test set); built a runnable multi-track video editor web application from only simple instructions (8,192 lines of code, 1,868 calls, completed in 11.5 hours).

  Speech Capabilities

  • V2.5-TTS Series: Comprehensive upgrades in speech synthesis naturalness and support for multiple languages/dialects/voices.
  • V2.5-ASR: Improved speech recognition accuracy and real-time performance, supporting multiple languages and dialects.

How to Use MiMo-V2.5

• Web Experience: Visit the MiMo Studio official website, register and log in, then select the MiMo-V2.5 model to start a conversation.
• API Development Integration: Visit the API open platformhttps://platform.xiaomimimo.com/#/token-plan注册开发者账号，创建应用获取 API Key。

Key Information and Usage Requirements of MiMo-V2.5

• Access Methods
  • MiMo Studio Experience：https://aistudio.xiaomimimo.com
  • API Calls：https://platform.xiaomimimo.com
• Open-source Plan: MiMo-V2.5-Pro and MiMo-V2.5 will soon be open-sourced globally.
• Pricing and Billing
  • MiMo-V2.5：1x Credits（1 Token = 1 Credit）
  • MiMo-V2.5-Pro：2x Credits（1 Token = 2 Credits）
• Context Window: The entire series supports an ultra-long context of 1M tokens (about 750,000 Chinese characters), with no additional charge for long document analysis.

Core Advantages of MiMo-V2.5

• Token Efficiency Revolution: When reaching the same score on the ClawEval benchmark, V2.5-Pro saves 42% of tokens compared to Kimi K2.6, and V2.5 saves 50% of tokens compared to Muse Spark, significantly reducing large-scale deployment costs.
• Long-horizon Task Stability: With an appropriate runtime framework, it can stably complete long-horizon tasks with nearly a thousand rounds of tool calls in a single session, and maintain logical consistency and self-correction over extremely long cycles (for example, in the compiler task, it self-diagnosed and recovered after refactoring in round 512).
• Native full-modality fusion: Text, image, audio, and video capabilities are integrated into a single model, with no need to switch between multimodal models and text models.
• Inference speed: V2.5 has an average inference speed of 100~150 tokens/s, while V2.5-Pro is 60~80 tokens/s; the base version is better suited for latency-sensitive scenarios.

MiMo-V2.5 project address

• Project official website：https://platform.xiaomimimo.com/docs/news/v2.5-news
• HuggingFace model repository：https://huggingface.co/collections/XiaomiMiMo/mimo-v25

Comparison of MiMo-V2.5 with similar competing products

Application scenarios of MiMo-V2.5

• Complex software development: Autonomously completes the from-scratch construction and debugging of large engineering projects such as compilers, video editors, and web applications, and can replace several days to several weeks of work by human experts.
• Multimodal content analysis: Upload a photo of a refrigerator to get recipe suggestions, parse video tutorials to generate step summaries, and record meetings to extract to-do items.
• Scientific research and engineering automation: Analog circuit EDA design, performance optimization, long-horizon analysis of academic literature, and cross-document reasoning.
• Intelligent customer service and interaction: In-vehicle, smart home, and real-time interactive systems based on full-modality perception (speech + images + text).
• Enterprise-level Agent deployment: In large-scale API call scenarios, high token efficiency enables cost-controllable automated workflows.