Lenovo has previewed the AI Host P7, a 300-gram, fanless desktop box that the company says can run language models of up to 125 billion parameters locally, without sending data to a cloud server. Lenovo showed the device on October 24 at CSDN's 2026 1024 Programmers' Day and AI Builders Conference and opened an early-access program aimed at developers and professional users.
The P7 is roughly the size of a coaster, draws at most 30 watts, and runs below 35 dBA by Lenovo's own figure. Lenovo has not announced a price or a general release date, and it has not said how many early-access places it is offering or what participants will pay.
Two domestic chips split the work. The Houmo M50 handles model inference at about 10 watts and is rated at 160 TOPS INT8 or 100 TFLOPS BF16. It carries 12 to 48GB of LPDDR5X memory on a 192-bit bus at up to 153.6 GB/s and connects to the rest of the system over four PCIe Gen4 lanes.
Key facts
- 01Lenovo previewed the AI Host P7 on October 24 at CSDN's 2026 1024 Programmers' Day and AI Builders Conference.
- 02Lenovo says the 300-gram, fanless box runs models up to 125B parameters with a 256K context while drawing at most 30W.
- 03A Houmo M50 inference chip rated at 160 TOPS INT8 pairs with a CIX P1 system chip for a quoted 190 TOPS.
- 04The advertised 80GB of memory is split across two separate pools on the two chips, not one unified pool.
- 05Lenovo has opened an early-access program for developers but has not announced a price or release date.
The CIX P1 runs the operating system and schedules work. It is a 6nm part with twelve Armv9.2 cores, a ten-core Immortalis-G720 GPU, and a 30 TOPS NPU, for 45 TOPS in total, with support for up to 64GB of shared LPDDR5 at about 100 GB/s. Zhang Hua, a Lenovo senior vice president, said at the product's May launch that CIX custom-built the processor for Lenovo. Lenovo's headline figure of 190 TOPS is the M50's 160 plus the P1's 30 TOPS NPU.
Lenovo's claims have grown since May. The original launch promised 122 billion parameters and a 128K-token context window; the October preview claims 125 billion parameters and a 256K window. CIX reportedly says a 122B model with a 128K context runs at 50 tokens per second, and Lenovo says its mixed INT4 and INT8 quantization averages 98.3% of a BF16 baseline across four benchmarks it did not name.
The software side is built around local-first use with a cloud fallback. The box ships with Lenovo's Tianxi AI preinstalled, and an auto mode routes requests between local and cloud models. It also includes a model manager, a hardware dashboard, a task mode for always-on agent jobs, and an API address that external tools can point to. The model library lists Qwen3.6-35B-A3B, Qwen3.8-Flash-Next, and a 27B-parameter model.
Sun Ke, who leads Lenovo's AI host line, framed the P7 around the cost of using AI rather than raw capability, arguing that people should not have to track the price of every token or hesitate before handing an AI an important task. That pitch positions the P7 as a replacement for paid cloud token usage rather than a faster PC. Its smaller sibling, the AI Host mini, is about half the size of a Mac mini, delivers 45 TOPS at 15 watts average power, and launched at 2,999 yuan, roughly $420.
The numbers so far come from Lenovo, Houmo, and CIX and have not been independently tested. Intelligent Living notes that the advertised 80GB of memory sits in two separate pools on two chips rather than one unified pool, and that the 50-tokens-per-second figure is plausible for a sparse mixture-of-experts model but not for a dense model of that size.
Lenovo first showed the P7 in May with a plan to open crowdfunding on July 1, which did not lead to a release. The early-access program opened at the October 24 preview is the next step, and Lenovo has still not given a general release date.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
