Lenovo showed the AI Host P7 preview on October 24, 2026, saying the palm-sized device, which weighs about 10.6 ounces (300 grams), could run local models with up to 125 billion parameters and a 256K-token context window. The company also announced an early-access program. The figures describe Lenovo’s preview claims, while the TOPS numbers refer to different parts of the hardware.
What the M50 and P1 TOPS figures describe
TOPS means trillions of operations per second. Lenovo’s event description rates the Houmo M50 inference chip at 160 TOPS at 10 watts. It describes the CIX P1 system-on-chip as having a 30-TOPS neural processing unit (NPU) and 45 TOPS of heterogeneous compute across its processing components.
A technical analysis by Intelligent Living explains the reported 190-TOPS total as the M50’s 160 TOPS plus the P1 NPU’s 30 TOPS. The P1’s 45-TOPS heterogeneous figure is a separate, broader total.
| Component or figure | TOPS | What it describes |
| Houmo M50 | 160 at 10 W | Inference chip |
| CIX P1 NPU | 30 | Neural processing unit |
| CIX P1 heterogeneous compute | 45 | P1’s broader compute total |
| Reported combined figure | 190 | M50 at 160 plus P1 NPU at 30 |
TOPS counts operations per second; tokens per second counts generated text units. The measures describe different things, so the compute figures do not state how quickly the P7 generates text.
How the reported 80 GB of memory is arranged
Technical reports by Intelligent Living and AI Chat Daily describe the P7’s reported 80 GB as separate memory pools attached to the M50 and P1, rather than one unified pool. Lenovo’s event description separately says the P1 supports up to 64 GB of LPDDR5 memory at 6,400 Mbps. That P1 figure describes the chip’s stated memory support, not the P7’s total installed memory.
The P1 is described as a 6-nanometer chip with a 12-core Armv9.2 CPU. Read together, the component figures distinguish the system’s two compute chips, the P1’s own memory support, and the separate pools described in technical reporting.
Models and software Lenovo named
Lenovo listed Qwen3.8-Flash-Next, Qwen3.8-27B, and Qwen3.6-35B-A3B among the local models, along with model categories for image design, document recognition, speech recognition, and speech synthesis.
The preview description says Tianxi AI comes preinstalled. Its Auto mode is designed to choose between local and cloud models according to task complexity. Token Manager configures an API or URL to connect local models with coding tools and agent applications.
Lenovo also said mixed INT4/INT8 quantization averaged 98.3% of the BF16 model score across four benchmarks. That figure is a benchmark score comparison, separate from the P7’s TOPS ratings.
Lenovo’s preview and early-access program
At the October 24 event, Lenovo announced early access for initial users to try the P7 and provide feedback. The company said general sales would follow completion of hardware and software development.
