On April 23, 2026, Xiaomi released the MiMo-V2.5 series of full-modal large models. A month later, the model’s Arena performance confirms Xiaomi’s investment in large models has entered the first tier.
Arena Performance
- Arena text leaderboard: global sixth (~1489 range), highest among open-source models
- Agent index: open-source #1, top five across all models globally
- Global open-source comprehensive intelligence index: tied #1
These results mean MiMo-V2.5-Pro’s real-user conversation performance has surpassed GPT-5.5 (Arena text seventh) and most closed-source models.
Core Capabilities
Agent capability. Agent index #1 among open-source indicates significant advantages in tool calling, multi-step task planning, and autonomous execution — aligned with Xiaomi’s IoT ecosystem experience.
Million-level long context. Supports 100K+ token context window, practical for codebase analysis, legal document review, and long video subtitle understanding. At equivalent benchmark scores, MiMo-V2.5-Pro’s token usage efficiency is relatively lower.
Full-modal coverage. The series covers text, speech (V2.5-TTS), and more — one of the few open-source model families covering multiple modalities.
Ecosystem Compatibility
Xiaomi announced MiMo-V2.5 series is compatible with nearly all Chinese inference chips — an important usability indicator for domestic enterprise users. The model can be deployed without relying on NVIDIA GPUs.
Comparison
| Dimension | MiMo-V2.5-Pro | Qwen3.6-35B-A3B | GLM-5.1 |
|---|---|---|---|
| Arena Text | Global #6 / OS #1 | Not Top 10 | Not Top 10 |
| Code | Moderate | Near Claude 4.5 | Arena Code #5 |
| Long Context | Million+ | Million+ | Unclear |
| Multi-modal | Text + Speech | Text-focused | Text-focused |
Recommendations
- IoT / Smart Home: MiMo-V2.5-Pro’s Agent capability and Xiaomi ecosystem integration.
- Chinese chip deployment: Broad compatibility reduces hardware dependency.
- Long document processing: Million-token context + efficient token usage.
Main sources: