Xiaomi Launches MiMo-V2.5 Series With 1M Context and Reasoning Tokens
Agentic Coding
LLM
AI Economics
AI Agent
Multimodal
Performance
Xiaomi, a consumer electronics and software company, launched the MiMo-V2.5 series on OpenRouter with native omnimodal capabilities for text, images, and video. Both variants feature a 1M token context window and support reasoning tokens, allowing the model to perform internal deliberation before generating a final response.
This release extends the availability of Xiaomi's agent-centric models as developers shift toward long-running autonomous workflows. By optimizing for agentic performance and coding stability, the series addresses reliability issues in multi-step tasks. It follows a trend of optimizing models specifically for agent pipelines.
You can integrate mimo-v2.5 into agent frameworks to handle massive codebases in a single pass. The model is available via the OpenRouter API at $0.40 per million input tokens and $2 per million output tokens. To use reasoning, enable the reasoning parameter and preserve the reasoning_details array.
Frequently asked questions
- What is the pricing for Xiaomi MiMo-V2.5 on OpenRouter?
- MiMo-V2.5 is priced at $0.40 per million input tokens and $2.00 per million output tokens on the OpenRouter platform. This pricing structure is designed to be highly token-efficient, offering pro-level performance for agentic tasks at approximately half the inference cost of comparable high-end models in the current market.
- How large is the context window for the MiMo-V2.5 series?
- Both MiMo-V2.5 and MiMo-V2.5-Pro feature a massive context window of 1,048,576 tokens. This 1M token capacity allows the models to process entire codebases, lengthy documents, and extended conversation histories in a single pass, which is particularly useful for complex agentic workflows that require maintaining a large amount of state.
- How do reasoning tokens work in the MiMo-V2.5 models?
- MiMo-V2.5 supports reasoning tokens, which are internal thinking tokens the model generates to work through logic before providing a final response. Users can enable this by using the reasoning parameter in their API request. Accessing the reasoning details array allows developers to see the step-by-step internal thinking process used by the model.
- What multimodal tasks can MiMo-V2.5 handle?
- MiMo-V2.5 is a native omnimodal model, meaning it was built to understand multiple types of data within a single architecture. It specifically shows improvements in multimodal perception for image and video understanding tasks, surpassing previous versions like MiMo-V2-Omni. This makes it suitable for agents that need to interpret visual information alongside text.
- What are the primary use cases for the MiMo-V2.5 series?
- The MiMo-V2.5 series is specifically optimized for long-running agent tasks and advanced coding. Its combination of a 1M context window, reasoning capabilities, and high token efficiency makes it an ideal engine for autonomous agent frameworks. It excels at tasks requiring complex instruction decomposition, stable tool use, and deep reasoning over large datasets.
