Vibeleaderboard
Xiaomi

ai lab

Xiaomi

Xiaomi MiMo matters because it joins open model development to one of the world's largest consumer-device ecosystems. The useful question is not whether a phone company has become a frontier lab overnight, but whether MiMo can turn credible language, multimodal, and embodied research into software that works across Xiaomi's phone, vehicle, and home products.3,4,7

Official site ↗Models in the Atlas →

Profile

Overview

A model program inside Xiaomi

Xiaomi MiMo is the foundation-model program inside Xiaomi, the Beijing consumer-electronics company founded by Lei Jun and partners in 2010. The lab is not a standalone company, and its releases should not be treated as the whole of Xiaomi's long AI history. MiMo became a distinct public model identity in 2025, when the Xiaomi LLM-Core team released a technical report and open checkpoints for a compact language model trained for mathematics, code, and general reasoning.1,2

From compact reasoning to multimodal systems

The first MiMo report described a 7-billion-parameter model trained on about 25 trillion tokens, followed by supervised fine-tuning and reinforcement learning. The program then widened beyond text. MiMo-VL applied supervised and reinforcement-learning stages to a 7-billion-parameter vision-language model, while MiMo-Embodied joined autonomous-driving and embodied-agent data in one cross-embodiment model. Those releases establish MiMo as a parent-company research and product line spanning language, vision, and physical environments, rather than a single chatbot.2,3,4

Sparse models and agents

MiMo's second generation moved toward sparse, agent-oriented systems. MiMo-V2-Flash used a mixture-of-experts architecture with 309 billion total parameters and 15 billion active parameters, plus a long context window and explicit training for reasoning, coding, and tool use. Xiaomi later connected the line to MiMo-V2-Pro and MiMo-V2.5 variants. Independent model trackers recorded competitive but uneven results, which is more useful evidence than repeating the company's comparisons against selected frontier systems.6,8,9

A route into devices

Xiaomi has a distribution path that most model laboratories do not: phones, vehicles, home devices, and the HyperOS software layer. Its 2026 interim reporting says the company used MiMo in Miloco 2.0, an open smart-home copilot, and cites heavy third-party API usage for MiMo-V2.5. That makes the program strategically important even when public model rankings are mixed. It also creates an attribution boundary: device sales and Xiaomi's broader robotics work are context for MiMo, not achievements of the model team unless a source directly connects them.7,5

Notable contributions

  1. 01Notable contribution: a documented compact reasoning recipeThe first MiMo release documented a coordinated pretraining and post-training recipe for a 7-billion-parameter reasoning model. It is a notable compact-model contribution, not a claim that Xiaomi invented reinforcement learning for reasoning.2
  2. 02Notable contribution: open cross-embodiment modelingMiMo-Embodied combined autonomous-driving and embodied-AI tasks in one openly released model and evaluation program, creating a concrete bridge between Xiaomi's vehicle and robotics research domains.4
  3. 03Notable contribution: home-specialized vision-language modelsMiMo-VL-Miloco adapted the broader vision-language line to home activities, gestures, and long-form video, and released checkpoints plus a home-scenario evaluation toolkit.5