MiMo-V2.5-Pro

Overview

MiMo-V2.5-Pro is an open-weight text model from Xiaomi MiMo that entered public beta on April 23, 2026. The higher-capacity member of the MiMo-V2.5 family uses a 1.02-trillion-parameter sparse Mixture-of-Experts (MoE) architecture with 42 billion active parameters and is designed for long-horizon agents, software engineering, and complex reasoning.

Key Features

• 1.02T Sparse MoE Architecture: Uses 1.02 trillion total parameters, 42 billion active parameters, 384 routed experts, and eight selected experts per token.
• Hybrid Long-Context Attention: Interleaves 60 Sliding Window Attention layers with 10 global-attention layers and a 128-token sliding window. Xiaomi reports that the 6:1 local-to-global pattern reduces KV-cache storage by nearly 7x.
• 1M Context and 128K Output: Supports a one-million-token context window and up to 128K output tokens for repository-scale work, long documents, and extended agent trajectories.
• Agent and Coding Evaluation: Xiaomi reports 1,581 Elo on GDPVal-AA, 63.8 on ClawEval, 57.2 on SWE-Bench Pro, 78.9 on SWE-Bench Verified, and 68.4 on Terminal-Bench 2.0.
• Controllable Reasoning and Tools: Deep thinking is enabled by default but can be disabled. The API supports streaming, function calling, and structured output.

Best Use Cases

• Long-Horizon Software Engineering: Repository navigation, multi-file implementation, debugging, testing, and tool-driven work that must remain coherent across long trajectories.
• Complex Agent Workflows: Tasks involving repeated function calls, planning, execution, and correction over large working contexts.
• Long-Document Reasoning: Analysis and synthesis across extensive code, technical documents, research material, or conversation history.
• Self-Hosted Model Development: Commercial inference, fine-tuning, or secondary training where MIT-licensed weights and framework-level deployment control are required.

Capabilities and Limitations

Capability	Description
Reasoning	Deep thinking is enabled by default and can be turned off with thinking.type. Xiaomi reports 48.0 on Humanity's Last Exam with tools and 34.0 without tools.
Coding	Xiaomi reports 57.2 on SWE-Bench Pro, 78.9 on SWE-Bench Verified, 68.4 on Terminal-Bench 2.0, and an implementation rank of 3.4 on FrontierSWE.
Creative Writing	Supports general and long-form text generation.
Multimodal	Text input and text output.
Context Window	1M tokens.
Max Output	128K tokens.
Tool Use	Supports function calling, structured output, streaming, and Xiaomi's web-search tool. In thinking-mode agent conversations, tool-call history must retain the complete reasoning_content field.
Multilingual	The official model repository identifies English and Chinese support. Xiaomi reports 83.6 on GlobalMMLU, 91.5 on C-Eval, and 90.2 on CMMLU for the Pro Base model, but does not publish a complete language-support list.

Known Limitations

• mimo-v2.5-pro is a text model and does not provide the full-modal understanding available in mimo-v2.5.
• In thinking mode, custom temperature and top_p values are ignored; the API forces 1.0 and 0.95. Multi-turn tool workflows that omit historical reasoning_content can fail with HTTP 400 or lose context quality.

Pricing

Model	Input (Credits/Token)	Cache Write (Credits/Token)	Cache Read (Credits/Token)	Output (Credits/Token)	Web Search (Credits/Use)
MiMo-V2.5-Pro	0.435	0.435	0.0036	0.87	-

Pricing note

Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through top-up bonuses and account benefits. Specific prices, bonus Credits, and account benefits are subject to the platform display and final billing records.