Xiaomi has gone from having no public AI models to building a sprawling portfolio that covers reasoning, vision, audio, voice cloning, coding, and autonomous phone control. The surprising part is not that the company joined the AI race, but how quickly it moved from hobbyist-sounding releases to Xiaomi AI models that are drawing real developer traffic and, in some cases, taking on bigger names head-on.
That matters because Xiaomi is not treating AI as a one-off feature add-on. It is tying the work to phones, smart home gear, wearables, cars, and its own operating system, which is a much more ambitious bet than simply shipping another chatbot app. If the hardware company can keep the models useful, it gets a shot at owning the whole stack.
MiMo models are Xiaomi’s public proof of concept
The headline act is MiMo-V2-Flash, a 309-billion-parameter model that only activates about 15 billion parameters at a time through a Mixture-of-Experts design. Xiaomi says it can generate 150 tokens per second, ranks among the top two open-source models on reasoning benchmarks, and matched GPT-5 and Claude 4.5 Sonnet on SWE-Bench Verified.
MiMo-V2-Pro pushed the idea further. It has over one trillion total parameters, 42 billion active parameters per pass, and a one-million-token context window. In plain English, it is built to swallow huge amounts of information and keep working through messy, multi-step jobs instead of pretending everything is a neat Q&A exchange.
- MiMo-V2-Flash: 309 billion total parameters, about 15 billion active, 150 tokens per second
- MiMo-V2-Pro: over one trillion total parameters, 42 billion active, one million-token context
- MiMo-V2.5-Pro: 1.02 trillion parameters, multimodal support for text, image, audio, and video
Xiaomi also built the models with developer bait that is hard to ignore: low pricing, free launch access in some cases, and open-source releases that make experimentation cheap. That is a familiar playbook in AI, but Xiaomi is using it with unusually aggressive speed. The company had already attracted a meaningful share of traffic on OpenRouter by early April 2026, which suggests people were actually using the models rather than just applauding them from afar.
Audio and voice are Xiaomi’s quieter advantage
Not all of Xiaomi’s AI effort is aimed at chat and code. MiDashengLM-7B, released in August 2025, was trained on 38,662 hours of audio and is designed to understand music, environmental noise, emotion, and other non-verbal cues that standard speech systems often ignore. It is already embedded in Xiaomi’s electric vehicles and smart home devices, which is exactly where audio intelligence starts to feel practical instead of decorative.
Then there is OmniVoice, a text-to-speech model that supports 646 languages and can clone a voice from just a few seconds of reference audio. Xiaomi says it can train on 100,000 hours of audio in a day and run inference at up to 40x real-time speed using PyTorch. That is a very loud statement from a company that, until recently, was better known for selling affordable phones than for building multilingual voice systems.
Xiaomi also paired MiMo-V2.5-TTS with an ASR system for bilingual recognition, making the voice stack more complete than the average ”AI assistant” marketing slide. Add MiMo-Audio, whose encoder later fed into MiMo-V2.5, and the pattern is obvious: Xiaomi is trying to make audio a core capability, not an accessory.
HyperAI, Xiao AI and miclaw pull the stack into products
On consumer devices, Xiaomi is putting the models to work through Xiao AI and HyperAI. Xiao AI has been upgraded with better memory, smarter home control and text-to-image generation, while HyperAI adds translation, writing help, speech summarization and photo editing inside HyperOS 2. For global devices, Xiaomi has also used Google Gemini as a backend, which is a polite way of saying the company is happy to borrow where it needs to.
The more interesting move is miclaw, Xiaomi’s autonomous AI agent currently in closed beta. It is not framed as a chatbot; it is meant to interpret a task and then carry it out by opening apps, navigating interfaces, filling forms and using system tools. That is the kind of feature that sounds mundane until it works, at which point every basic phone workflow suddenly feels old.
Privacy is the obvious pressure point, and Xiaomi is trying to pre-empt the criticism by saying miclaw does not use user interactions to train models. Whether that reassurance sticks will depend less on slogans and more on how comfortable people are letting software act on their behalf across phones, watches, cars, and home devices.
Xiaomi’s AI spending is now impossible to ignore
Lei Jun has said Xiaomi will invest at least $8.7 billion in AI over the next three years, with annual R&D spend projected to reach around 40 billion yuan ($5.7 billion) in 2026. That is serious money, but it also reflects a broader truth: Xiaomi is no longer dabbling in AI features around the edges of its business.
The company is aiming for a ”grand convergence” of its own chip, OS and AI model in a single device, and that is where this starts to look less like model bragging and more like platform strategy. If miclaw and HyperOS 4 make the system genuinely useful, Xiaomi could become one of the few device makers with both the hardware and the AI layer under its own control. If not, it will still have shipped a lot of impressive code. But impressive code is not the same as daily habit, and that is the part Xiaomi has to win next.