Lenovo AI Host P7: what's confirmed about the 300-gram box said to run 125B-parameter models locally

The Lenovo AI Host P7 is a palm-sized AI computer that Lenovo unveiled in Beijing on May 19, 2026 and showed again, as a preview edition for developers, in China on October 24, 2026. It's pitched at people who want to run large language models locally instead of paying per token. Lenovo's chip partner Houmo says it weighs about 300 grams, and Chinese reports say it runs models of up to 125 billion parameters, which would put it in territory usually reserved for much larger machines. But the details come from Chinese tech press and Lenovo's chip partners, not from a Lenovo spec page, and the reported model size has changed between announcements. There's no price or release date. Running small models on your own hardware comes down to memory, as our guide to choosing a model for your laptop explains.

This story is based on Houmo's May 2026 announcement of the P7 launch, reports by Chinese tech sites Kuai Keji (MyDrivers) on October 18 and October 24, 2026 and IT Home on October 24, a post by chipmaker Cixin's official Weibo account, Houmo's M50 product page, and reporting by TechRadar and Liliputing on Lenovo's cheaper AI Host mini PC, checked October 27, 2026. No Lenovo announcement or product page for the P7 was found.

The short version

• What: a preview of a small AI computer for local models, with an early-access program. Not on sale yet.
• Where and when: unveiled at a Lenovo event in Beijing on May 19, 2026; a preview edition was shown on October 24, 2026 at CSDN's 1024 Programmer's Day and AI Builders Conference in Changsha, China.
• Chips: Cixin's P1 processor plus Houmo's M50 AI accelerator, 190 TOPS combined, per Houmo.
• Model size: 122B parameters in Houmo's May announcement and an October 18 report; up to 125B in reports from the October 24 event.
• Price and availability: not announced. Houmo said in May that crowdfunding would open July 1, but no campaign was found; on October 24 Lenovo said only that sales would start soon, according to Kuai Keji.

What is the Lenovo AI Host P7?

Lenovo first unveiled the P7 at a Tianxi AI launch event in Beijing on May 19, 2026, according to an announcement by Houmo, which makes its M50 accelerator. Houmo described it as an AI-native personal and home edge device built for agents, and said crowdfunding would open on July 1. No report of that crowdfunding campaign was found, and the P7 still had no price or sale date in October.

According to Kuai Keji's October 24 report, Lenovo presented the AI Host P7 as an "AI productivity hub" for programming and office work that lets developers run large models locally and connect them to coding tools and agent apps through local API endpoints. It runs Lenovo's Tianxi AI assistant, which can pick between local and cloud models. Sun Ke, who heads the AI Host product at Lenovo, said, as quoted by Kuai Keji (in translation): "Real productivity shouldn't require calculating token costs before every deep use."

Cixin, which makes the P7's main processor, confirmed the preview and the early-access program in a post on its official Weibo account on October 25. It described the P7 as able to run "hundred-billion-parameter-class" models locally, but gave no exact size, weight or price.

The P7 shouldn't be confused with Lenovo's cheaper AI Host, a mini PC that went on pre-order in China in June 2026 for 2,999 yuan with an 8GB Cix CD8180 chip, as TechRadar and Liliputing reported. The P7 is a different, more powerful design.

What are the AI Host P7's specs?

These figures come from the chip vendors and Kuai Keji's reports, not from a Lenovo spec sheet.

Last verified: October 27, 2026

Houmo's May announcement of the P7 lists 190 TOPS of local AI compute, local models of up to 122B parameters with a 128K context, up to 50 tokens per second running locally, up to 80GB of memory, 30W maximum power, under 35 dB(A) of noise and a weight of 300 grams. Houmo's own page for the M50 lists 160 TOPS at INT8, 100 TFLOPS at BF16, about 10W typical power, 12GB to 48GB of LPDDR5 or LPDDR5X memory with up to 153.6 GB/s of bandwidth, and says a single M50 can run models of up to 120 billion parameters. Kuai Keji reported that Lenovo uses mixed INT4 and INT8 precision and claims an average of 98.3% of full BF16 accuracy across four benchmarks it didn't name, a Lenovo claim with no published method. Kuai Keji also reported that the P1 is a 6nm chip with a 12-core Armv9.2 CPU and a 30 TOPS NPU, 45 TOPS in total, and, in its October 18 report, that the P7 draws 30W at most and runs below 35 dB(A).

TOPS figures say little on their own about how fast a model runs, as our NPU explainer details. For large language models, memory size and bandwidth usually matter more.

Can it really run a 125B model?

That's the part to treat with care. Houmo's May announcement and Kuai Keji's October 18 preview of the October event both said the P7 could run a 122B-parameter model with a 128K-token context at up to 50 tokens per second. Kuai Keji's October 24 report from the event, and IT Home's, said up to 125B parameters with a 256K context, and 122B with 128K "with expansion," still at up to 50 tokens per second. Neither report names the 125B model, and Lenovo hasn't published benchmark conditions.

A speed like 50 tokens per second is far more plausible for a mixture-of-experts model, which uses only a fraction of its parameters for each token, than for a dense model of that size. The local models Kuai Keji listed for the P7 include Qwen3.6-35B-A3B, a mixture-of-experts model whose "A3B" suffix denotes about 3 billion active parameters, and Qwen3.8-27B. Lenovo hasn't said which architecture its large-model figures refer to. If you want to run open models on hardware you already own, our Ollama guide covers the steps.

How much does the Lenovo AI Host P7 cost, and when can I buy it?

Lenovo hasn't announced a price or release date. Houmo said in May 2026 that crowdfunding would open on July 1, but no report of that campaign was found. Kuai Keji reported on October 24 that the P7 will go on sale "soon" and that Lenovo opened an early-access program at the event to recruit first users. None of the reports says whether the P7 will be sold outside China, and the cheaper AI Host mini PC was China-only.

What's confirmed, reported and uncertain

• Confirmed (chip vendors): the May 19, 2026 launch in Beijing, the planned July 1 crowdfunding, 190 TOPS, about 300 grams, up to 80GB of memory, 30W, 122B parameters with 128K context and up to 50 tokens per second (Houmo's announcement); the October 24 preview at CSDN's 1024 conference and the early-access program (Cixin's official Weibo); the M50's specs and its 120B-parameter rating (Houmo).
• Reported (Kuai Keji, IT Home): 64GB of memory on the P1, the 125B figure with 256K context, the 98.3% accuracy claim, and Sun Ke's quote.
• Uncertain: the exact largest model and its architecture, whether the July 1 crowdfunding happened, the price, the sale date and whether the P7 will sell outside China. No Lenovo product page or English-language report from a major outlet was found.