Lenovo introduced the AI Host P7, a device it developed for the personal AI market. The coaster-sized device weighs 300 grams and can run models with up to 125 billion parameters directly on the desk without an internet connection.
This scale normally requires a workstation with tens of gigabytes of VRAM and thousands of watts of power. Lenovo rebuilt the architecture from the ground up to fit the same workload into a small box.
What’s inside the 300-gram body?
At the heart of the device is Houmo AI’s M50 compute-in-memory (CIM) chip. In this approach, data is not constantly moved between memory and the processor; matrix multiplications are performed directly inside memory cells. This eliminates the von Neumann bottleneck and significantly reduces power consumption.
The second component is CIX Technology’s P1 agentic SoC. Task planning, tool use, and multi-step workflows run through this unit. The third step is dynamic mixed quantization: the model’s critical reasoning layers are kept at 8-bit and attention layers at 4-bit, reducing the memory footprint while preserving accuracy.
The real claim is data sovereignty
The device targets users who don’t want to send data to cloud tools: law firms, financial advisors, healthcare organizations, and software teams. Every document processed on the P7 and every line of code written stays on local hardware.
Connected to the local network, the device serves phones, tablets, and laptops like a local API endpoint. Lenovo says the device draws as much power as a phone charger and runs completely silently with passive aluminum fins.
No price or date yet
The launch did not reveal the device’s price or release date. There is also no independent measurement of how many tokens per second it generates on 125-billion-parameter models. Until those two data points arrive, it is hard to say how well the P7 delivers on its promise.
Still, the direction is clear: AI will run not only in massive data centers, but also in small boxes on the desk.
Source: WebTuring
Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.