Xiaomi MiMo-V2.5 Dual Models Open-Sourced: 1T MoE + 310B MoE, Million-Token Context, 100T Token Incentive Program

Xiaomi MiMo-V2.5 Dual Models Open-Sourced: 1T MoE + 310B MoE, Million-Token Context, 100T Token Incentive Program

Key Takeaways

Xiaomi open-sourced two large language models in late April 2026, using MoE architecture spanning 1T and 310B parameter scales, both supporting million-token context windows. The simultaneous MiMo Orbit developer incentive program—up to 1.6 billion free tokens—directly competes with domestic vendors’ developer subsidy strategies.

Model Specifications & Architecture

DimensionMiMo-V2.5-ProMiMo-V2.5
Total Parameters1T310B
Active Parameters42B15B
Context Window1M Tokens1M Tokens
ArchitectureMoEMoE
LicenseMITMIT
PositioningComplex Agent + SEMultimodal Agent
Commercial Use✅ No extra license✅ No extra license

Three-pillar architecture supporting trillion-parameter sparse + million-length context:

  1. Hybrid Attention: Combines sliding window and global attention for efficiency at million-token scale
  2. Sparse MoE Routing: Only 42B of 1T total parameters activated, keeping inference costs manageable
  3. Long-Context Optimization: KV Cache management and attention decay specifically optimized for 1M token scenarios

Comparison with Competing Open Models

ModelTotal ParamsActive ParamsContextLicense
MiMo-V2.5-Pro1T42B1MMIT
Kimi K2.61T32B1MOpen Source
DeepSeek-V41.6T49B-Open Source
Qwen 3.6Various--Apache 2.0

MiMo-V2.5-Pro’s active parameters (42B) are close to Kimi K2.6 (32B), with comparable total parameters. On the Intelligence Index, MiMo V2.5 Pro scores ~54, behind Kimi K2.6 but the gap is narrow, both trailing GPT-5.5 (60 points).

100T Token Incentive: Competing for Developer Ecosystem

Xiaomi’s MiMo Orbit developer incentive program offers free tokens to global AI developers:

  • Maximum quota: 1.6 billion tokens
  • Review mechanism: Automatic review based on GitHub activity and AI usage history
  • Approval speed: Users report ~1 minute approval time
  • Target audience: High-quality AI application developers

This strategy mirrors Baidu and Moonshot’s developer subsidies—exchanging free compute for ecosystem lock-in and model feedback.

Luo Fuli’s Leadership: From DeepSeek to Xiaomi

The MiMo series is led by Luo Fuli (former Alibaba DAMO Academy, DeepSeek core member). In a 3.5-hour technical interview, she revealed:

  • Pre-train gap nearly closed: Domestic top teams are rapidly closing the gap with Anthropic in pre-training
  • Competition shifting to Agent RL: Next-gen model capabilities hinge on Agent reinforcement learning, not just pre-training scale
  • Open source is essential: Rapid community feedback and real-world data acquisition through open source

Action Recommendations

ScenarioRecommendationRationale
Local Agent deploymentMiMo-V2.5 (15B active)Low active parameters, reduced VRAM needs
Complex coding tasksMiMo-V2.5-ProDesigned for software engineering, 1M context
Commercial applicationsEitherMIT license, no extra authorization
Developer testingMiMo Orbit free quotaZero-cost model evaluation

MiMo-V2.5’s significance extends beyond parameters—Xiaomi enters open-source LLM competition as a hardware manufacturer. Combined with Xiaomi’s hardware ecosystem (phones, cars, IoT), MiMo has unique edge-cloud synergy potential.