Lenovo Debuts AI Host P7 Preview for Local 100B-Level Model Inference
On October 24, at the 2026 CSDN 1024 Programmer's Day and AI Builder Conference, Lenovo unveiled the AI Host P7 preview, an AI productivity hub designed for programming and office work. The device features a dual-chip design using domestic Chinese chips and supports running 100-billion-parameter large models locally.
According to Lenovo, the AI Host P7 comes with a range of local models and a token manager that lets users connect local models to coding tools and agent applications via one-click API/URL configuration, helping reduce cloud token costs. It is designed to run tasks continuously 7×24 hours while keeping important data on-device, with cross-device access through phones, tablets, and PCs.
The dual-chip solution consists of the M50 chip from Houmo Intelligent and the P1 agent SoC from Cixin. The M50 provides 160 TOPS at 10W power consumption based on a computing-in-memory architecture, while the P1 is built on a 6nm process with a 12-core Armv9.2 CPU and a 30 TOPS NPU, delivering 45 TOPS of heterogeneous computing power. The system also supports external AI accelerators. Lenovo introduced the “Tianxing AI Computing Architecture” to coordinate model optimization, heterogeneous resource scheduling, and inference engines for stable local operation of large models.
Sun Ke, head of Lenovo AI Host products, said true productivity should not require users to calculate token costs before every deep interaction. Lenovo previously published a white paper with IDC in September outlining the concept of “AI-native devices built for large models and agents,” and the P7 preview is positioned as the realization of that idea.
Lenovo has launched a “First-Experience Plan” to invite early users to test the AI Host P7 in real programming and office scenarios and participate in product co-creation before the official sale, which is expected soon.
