Xiaomi Launches MiMo-V2.5-Pro with 1M Context Window for Complex Agentic Tasks

Xiaomi released MiMo-V2.5-Pro on April 22, 2026, its flagship model featuring a 1,048,576 token context window and pricing at $1 per million input tokens and $3 per million output tokens.

Technical Specifications

The model's extended context window — just over 1 million tokens — positions it for integration with agent frameworks requiring long-form context retention. Xiaomi claims the model can autonomously complete professional tasks that would take human experts days or weeks, involving more than a thousand tool calls per task.

Benchmark Performance

According to Xiaomi, MiMo-V2.5-Pro achieves top rankings on:

• ClawEval: Benchmark for agentic capabilities (specific score not disclosed)
• GDPVal: General development proficiency evaluation (specific score not disclosed)
• SWE-bench Pro: Software engineering benchmark (specific score not disclosed)

Xiaomi has not released exact numerical scores for these benchmarks at launch.

Target Use Cases

The company positions MiMo-V2.5-Pro for three primary applications:

1. General agentic capabilities with extended autonomous operation
2. Complex software engineering tasks
3. Long-horizon tasks requiring persistent context across multiple steps

The model is available through OpenRouter, which routes requests to providers with automatic fallbacks for uptime optimization.

Pricing and Availability

MiMo-V2.5-Pro is now available at:

• Input: $1.00 per million tokens
• Output: $3.00 per million tokens

This pricing places it in the mid-tier range for flagship models, notably below comparable offerings from Anthropic and OpenAI with similar context windows.

What This Means

Xiaomi's entry with a 1M context window model at competitive pricing adds another option in the expanding agentic AI market. The emphasis on "thousands of tool calls" suggests optimization for complex multi-step workflows rather than single-turn generation. However, without published benchmark scores or independent verification, the claimed performance advantages over existing models remain unconfirmed. The model's actual differentiation will depend on real-world testing in software engineering and agent deployment scenarios.