Xiaomi MiMo V2.5 Pro

Xiaomi MiMo V2.5 Pro is Xiaomi's flagship agent-native large language model, built on a sparse Mixture-of-Experts (MoE) architecture with 1.02 trillion total parameters and 42 billion active parameters, released in August 2026. Designed for complex software engineering, long-horizon agentic workflows, and deep reasoning, the model supports a 1 million token context window (128K max output), employs hybrid sliding-window and global attention (6:1 ratio) to reduce KV-cache overhead, and includes three Multi-Token Prediction (MTP) modules to boost decoding throughput. The model ranks at the top on SWE-bench Verified, ClawEval, and GDPVal benchmarks, capable of autonomously completing professional tasks involving thousands of tool calls.

The iCreat platform offers this model via the endpoint https://api.icreat.ai/llm/openai/v1/chat/completions (OpenAI Chat Completions compatible), at $0.435 per million input tokens, $0.87 per million output tokens, and $0.004 per million cached read tokens.

Model Positioning

Xiaomi MiMo V2.5 Pro is positioned as Xiaomi's flagship agent-native LLM, designed for complex software engineering and long-horizon autonomous agentic workflows.

The model uses a 1.02T total / 42B active sparse MoE architecture with 384 routed experts (8 activated per token) across 70 transformer layers (1 dense + 69 MoE). Pretrained on approximately 27 trillion tokens, it supports a native 32K sequence length extended to 1M context. The model is text-only; the sibling MiMo V2.5 standard version supports text and image input.

On the iCreat platform, the model is served through an OpenAI Chat Completions compatible endpoint, and existing OpenAI SDKs work directly.

Core Capabilities

Trillion-Parameter Sparse MoE Architecture

1.02T total / 42B active parameters, 384 routed experts (8 per token), delivering trillion-scale representational capacity at 42B inference cost.

1 Million Token Ultra-Long Context

Supports 1,048,576 token context window, enabling ingestion of entire codebases, complete document sets, and long-running agent histories without repeated context compression.

Hybrid Attention Mechanism

6:1 ratio of Sliding Window Attention (SWA) to Global Attention with a 128-token sliding window, reducing KV-cache storage by nearly 7x while preserving long-context performance.

Multi-Token Prediction (MTP)​

Three lightweight MTP modules predict multiple future tokens during inference, tripling output speed under the described deployment configuration.

Agentic Reinforcement Learning Training

Post-training pipeline includes supervised fine-tuning, large-scale agentic reinforcement learning, and Multi-Teacher On-Policy Distillation (MOPD), explicitly targeting complex agent behavior rather than static Q&A.

Thinking Mode and Reasoning Effort

Supports thinking mode and reasoning_effort parameter (low/medium/high/xhigh/max), flexibly switching from rapid responses to deep reasoning.

Pricing

Note: The above are iCreat platform prices, consistent with Xiaomi's official overseas pricing (input $0.435, output $0.87, cache hit $0.0036). Cache write is free for a limited time. The iCreat platform price shall prevail.

Application Scenarios

• Complex software engineering and repository-scale development tasks
• Long-horizon autonomous agent workflows (thousands of tool calls)
• Large-scale codebase analysis and review
• Deep reasoning and complex mathematical problem solving
• Enterprise-grade document analysis and long-text processing

Model Comparison

Comparison Table 1: MiMo V2.5 Series

Note: Pro is a trillion-parameter MoE with stronger agent and coding capabilities; the standard version supports image input at a lower price.

Comparison Table 2: Same-Tier Agent Models

Note: MiMo V2.5 Pro blended pricing (input+output) is $1.305/M tokens, lower than GPT-6 Astra Economy ($6) and Kimi k1.5 ($2.65); it is the only MIT-licensed model in this comparison.

Why Choose Xiaomi MiMo V2.5 Pro?

• Trillion-parameter MoE architecture with 42B active cost for trillion-scale capacity
• 1 million token ultra-long context for entire codebase ingestion
• Hybrid attention reducing KV-cache by nearly 7x for efficient long-context inference
• Multi-Token Prediction modules tripling output speed
• Agentic RL training designed for complex software engineering and long-horizon tasks
• MIT open-source license for self-hosting (SGLang / vLLM)

Specifications

Architecture

The iCreat platform's Xiaomi MiMo V2.5 Pro is served through an OpenAI Chat Completions compatible endpoint, allowing users to call it with any OpenAI API-compatible SDK. Requests must include the API Key in the Authorization header. The request body contains the model field (xiaomi/mimo-v2.5-pro), a messages array (system/user messages), and optional thinking object (to enable thinking mode), reasoning_effort field (reasoning intensity), and stream parameter (streaming output). The model uses a sparse MoE architecture, activating only 42B parameters per token (out of 1.02T total), with hybrid sliding-window and global attention (6:1 ratio) reducing KV-cache overhead by nearly 7x, and three MTP modules for accelerated decoding. The model supports prompt caching to reduce repeated input costs, with cached read pricing at approximately one-hundredth of the standard input rate.

Notes

• The model supports text input and output only; for image input, use the sibling MiMo V2.5 standard version
• API Key should be passed via environment variables; avoid hardcoding it in source code
• The thinking mode and reasoning_effort parameters are optional; omit them for simple tasks to reduce latency
• Cache write is free for a limited time; cached read costs only $0.004/M tokens, significantly reducing costs for repeated queries
• Model weights are open-sourced under the MIT license; self-hosting requires FP8-capable hardware (e.g., B200_X8 or B300_X8)
• Please safeguard your API Key and avoid hardcoding it in client-side code or public repositories

Frequently Asked Questions

How is the billing calculated?​

Billing is based on actual token usage. Input is $0.435/M tokens, output is $0.87/M tokens, and cached read is $0.004/M tokens. For example, 1M input + 0.5M output costs $0.435 + $0.435 = $0.87. Cache write is free for a limited time. The iCreat platform price shall prevail.

What is the difference between MiMo V2.5 Pro and the standard V2.5?​

Pro is a trillion-parameter MoE (1.02T total / 42B active) designed for complex agent and software engineering tasks, supporting text input only. The standard version supports text and image input with a smaller parameter count and lower pricing (input $0.14, output $0.28). Choose Pro for agent coding; choose standard for image understanding or cost sensitivity.

How large is the context window?​

Supports 1 million tokens (1,048,576 tokens) context window with 128K (131,072 tokens) max output. Entire codebases or complete document sets can be ingested in a single request without segmentation.

Is thinking mode supported?​

Yes. Add "thinking": {"type": "enabled"} to the request body. In thinking mode, the model performs more internal reasoning before generating a response, suitable for complex tasks. Combined with the reasoning_effort parameter, you can adjust reasoning intensity (low/medium/high/xhigh/max).

Is the model open source?​

Yes. Model weights are open-sourced under the MIT license, including instruct and base checkpoints, and can be self-hosted via SGLang or vLLM. Deployment requires FP8-capable hardware (e.g., B200_X8 or B300_X8).

Is function calling supported?​

Yes. The model is compatible with the OpenAI Function Calling protocol, and also supports structured output and prompt caching. Web search is not natively supported (use external search tools via function calling).