Xiaomi’s Xuanjie O100 Prototype Is a Phone-Sized Demonstration of High-Speed Local AI

Xiaomi president Lu Weibing’s August 25, 2026 reveal was less a phone launch than a glimpse at what a local-AI-first handset might require. The Xuanjie O100 prototype combines the Xuanjie O3 smartphone processor with the O100 AI accelerator, removes the traditional rear-camera module, uses a rebuilt motherboard and adds active cooling rated for up to 10 watts. Running Xiaomi’s MiMo on-device model, it reportedly reached 295 tokens per second in testing. 37

That figure comes with an important qualification: the device is a technology-validation prototype, not an announced consumer product. 3 Its value is therefore not whether buyers can purchase it today, but what it demonstrates about the hardware trade-offs involved in running large models locally.

What Xiaomi revealed about the Xuanjie O100 prototype

The O100 prototype is designed around one objective: sustained, high-speed inference at the edge. Rather than treating AI as another feature inside a conventional camera-focused smartphone, Xiaomi paired two chips for collaborative computing:

• Xuanjie O3: the phone-class system-on-chip.
• Xuanjie O100: a dedicated high-bandwidth AI acceleration chip.

To accommodate that design, Xiaomi removed the conventional imaging module and redesigned the motherboard. It also added active air cooling capable of supporting up to 10 watts of cooling power. 34

The result is a phone-derived form factor built to showcase local large-model performance rather than normal handset priorities such as cameras, silent operation or maximum thinness. Xiaomi’s reported real-world result was 295 tokens per second while running its MiMo edge model. 35

That number should be read as a Xiaomi-reported result for a specific prototype and model, not as a universal benchmark for every AI model or future O100 device. The available reports do not establish how third-party models would perform under the same conditions.

O100 prototype vs. Xiaomi AI Cube

The two prototypes demonstrate different endpoints for Xiaomi’s local-AI strategy. The O100 is a mobile proof of concept; AI Cube is a stationary personal AI-computing terminal.

Feature	Xuanjie O100 prototype	Xiaomi AI Cube prototype
Form factor	Phone-derived AI demonstration terminal	Mini-PC and personal AI workstation
Chip configuration	Xuanjie O3 + O100	Xuanjie O3 + O100 + D100
Cooling or sustained output	Active cooling rated up to 10W	Claimed 150W sustained performance
Memory	Not specified in the cited reports	80GB unified memory
Demonstrated model capability	Xiaomi MiMo at a reported 295 tokens per second	120B model plus 3B on-device model, with fast/slow system switching
Product status	Prototype	Prototype

AI Cube has substantially more room for power delivery, memory and thermal management. Its aerospace-aluminum unibody includes 33,874 CNC-machined ventilation holes, and Xiaomi says the system can sustain up to 150 watts. Inside are all three Xuanjie chips—O3, O100 and D100—along with 80GB of unified memory. 4610

Xiaomi says AI Cube can deploy a 120-billion-parameter model alongside a 3-billion-parameter on-device model and switch between faster and slower systems depending on the task. The demonstrations were positioned around demanding workloads such as front-end development and complex programming. 37

In practical terms, AI Cube is closer to a compact local-AI workstation. The O100 prototype is more interesting as a test of whether that kind of inference acceleration can eventually be adapted to a consumer phone’s much tighter power and thermal envelope.

What the prototypes say about Xiaomi’s Xuanjie strategy

The prototypes fit into a broader three-chip product architecture Xiaomi presented at its August 24 Xuanjie technology briefing:

• O3 is the flagship smartphone SoC.
• O100 is the high-bandwidth accelerator for edge-side large-model computing.
• D100 is aimed at high-compute intelligent-driving workloads. 2129

This division suggests that Xiaomi is building an AI-computing stack across phones, personal computers and vehicles rather than relying only on cloud processing. That is an interpretation of the hardware direction, not a confirmed promise that the two prototypes will become retail products.

The commercial timeline also matters. Xiaomi says the O3 has entered mass production and is scheduled to debut in the Xiaomi 18 Fold in September. The O100 and D100 have been developed but are planned for commercialization later, with Xiaomi’s community announcement describing that timing as the following year. 1829

The key takeaway: a demonstration of trade-offs, not a new phone

The Xuanjie O100 prototype makes Xiaomi’s local-AI ambitions tangible, but it also exposes the cost of chasing sustained inference speed. A conventional camera module had to go, the motherboard had to be rebuilt and active cooling had to be added. That is a very different design priority from a mainstream flagship phone.

AI Cube shows the other side of the equation: give the system a larger chassis, far more power and more memory, and Xiaomi can demonstrate much larger local models and heavier workloads. Together, the devices show a company testing multiple form factors for self-developed silicon and on-device AI.

For now, the most concrete product takeaway is the O3-powered Xiaomi 18 Fold. The O100 prototype is better understood as a research and engineering statement: Xiaomi believes fast local large-model inference may eventually belong in phones, but the path to a practical consumer device still requires solving cooling, power, software compatibility and product-design constraints.