Lenovo AI Host P7 Runs 125B Models Locally in a 300-Gram Box

Lenovo says its AI Host P7 will run language models with as many as 125 billion parameters without sending a token to a data center. It weighs about 300 grams and is roughly the size of a drinks coaster. On October 24, at CSDN’s 2026 1024 Programmers’ Day and AI Builders Conference, Lenovo showed a preview version and opened an early-access program that puts it in developers’ hands before general sale.

Lenovo first showed the P7 in May, when crowdfunding was to open on July 1. Five months later there is still no price and no release date. Most coverage has focused on the parameter count. The numbers that will decide whether the P7 is useful are 153.6 gigabytes per second, the memory bandwidth of the chip at its center, and 80 gigabytes, a memory figure that turns out to be two pools rather than one.

What Lenovo Announced and What It Left Out

The October briefing added what May’s concept stage lacked: the chips, the software stack, and a local model library. The early-access program, which Lenovo calls a first-taste plan, targets developers and professional users rather than the public. The company has not said how many places it holds or what participants pay.

Sun Ke, who leads the AI host line, framed the pitch around cost rather than capability. Real productivity, he said, should not begin with calculating the price of every token, nor with hesitating before handing over an important task. Lenovo is not selling a faster computer. It is selling an end-to-end meter running in the background.

The framing has institutional backing. On September 10, Lenovo and IDC published a white paper proposing a category they call the AI-native device, built for large models and agents rather than retrofitted for them. The P7 is the first retail expression of that argument, which makes the company asserting the category also the company grading it.

Every performance figure here comes from Lenovo, Houmo AI, CIX Technology, or CIX’s documentation. None has been independently tested. One clarification for search traffic: this is not the ThinkStation P7 workstation, and it has nothing to do with Lenovo AI Now, the assistant shipped on its laptops.

Two Chinese Chips Doing Two Very Different Jobs

Both suppliers are domestic, and neither is a household name outside China. Lenovo’s venture arm has invested in both and describes the relationships as spanning capital, technology, research, and product.

Houmo M50: Inference at 10 Watts

The M50 runs the models. Houmo AI, founded in 2020, builds it on a compute-in-memory architecture that places arithmetic units inside the memory array, so weights do not cross a bus for every operation. That is the bottleneck that makes ordinary accelerators so power-hungry during inference, and removing it is how the chip claims 160 TOPS at INT8, or 100 TFLOPS at BF16, on about 10 watts.

The rest of the specification matters more than the TOPS figure. According to Houmo’s own specification sheet, the M50 carries 12 to 48 gigabytes of LPDDR5X on a 192-bit bus at up to 153.6 gigabytes per second and connects to the host over four lanes of PCIe Gen4.

CIX P1: The Chip Running the Machine

The P1 is the computer. CIX Technology builds it on a 6-nanometer process with twelve Armv9.2 cores, a ten-core Immortalis-G720 GPU, and a 30 TOPS NPU, which CIX totals at 45 TOPS of heterogeneous compute. It supports up to 64 gigabytes of shared LPDDR5 at about 100 gigabytes per second. CIX’s NeuralOne software stack reaches developers through PyTorch, TensorFlow, and ONNX.

The chip in the P7 is not the standard part. Zhang Hua, a Lenovo senior vice president, said at the May launch that the P7’s main processor was custom-made for the company by CIX. The division of labor is unusually explicit: the P1 runs the system and schedules work rather than performing inference itself.

Why 190 TOPS Is Not 160 Plus 45

Lenovo quotes the P7 at 190 TOPS. Add the M50’s 160 to the P1’s 30 TOPS NPU, and you get exactly 190, which means the headline excludes the Arm cores and Immortalis graphics that form the rest of CIX’s 45 TOPS aggregate. Both figures are accurate against different denominators, and stacking them side by side overstates the machine considerably.

A larger caveat applies to every local AI box. TOPS describes prompt processing and quantized arithmetic throughput. It says little about how fast text appears, because generation streams each weight it needs through memory once per token. That is why a desktop with a huge TOPS rating can feel slower than a laptop, a point worth revisiting in any comparison of the 192GB AI mini PCs now reaching reviewers.

The 80GB Is Two Pools, Not One

Lenovo’s other headline, up to 80 gigabytes of memory, looks like the unified pools that make Apple’s and AMD’s machines so effective. It is not. The M50 tops out at 48 gigabytes of its own LPDDR5X. The P1 supports up to 64 gigabytes of shared LPDDR5. The 80 gigabytes is the sum of two address spaces belonging to two different chips.

Memory inside the M50 runs at 153.6 gigabytes per second. Memory attached to the P1 runs at roughly 100. Traffic between them crosses four lanes of PCIe Gen4, which tops out near 8 gigabytes per second in theory and less in practice, an order of magnitude below either chip’s internal bandwidth.

A model small enough to sit inside one pool runs at that pool’s speed. A model that spans both is throttled by the slowest link, and Lenovo’s flagship claim requires exactly that: a 122-billion-parameter model at 4-bit occupies around 70 gigabytes, which cannot fit in the M50’s 48.

No coverage of the P7 has addressed this, and Lenovo has published no measured throughput for a model that spans both pools. It is the most important unknown in the specification sheet.

What 125B Parameters Actually Means for Speed

Lenovo’s own numbers have drifted upward since May. The original launch promised 122 billion parameters and a 128K context window. The October preview claims 125 billion and 256K. That is either a year of optimization or a rounding decision, and Lenovo has not said which.

Houmo’s individual claim sits awkwardly against its own hardware. A single M50 is rated for models up to 120 billion parameters but carries at most 48 gigabytes. A dense model of that size needs about 60 gigabytes at 4-bit, closer to 70 with file overhead. The figures reconcile only if the model is sparse, the quantization drops below 4 bits, or weights stream from system memory over a link too slow for useful speeds.

Then comes the arithmetic. At 4-bit, a model needs roughly 0.58 bytes per parameter. Multiply by parameter count, divide by bandwidth, and you get a hard ceiling on token generation before efficiency losses.

Read the table again, and the claim resolves. CIX says the platform runs a 122-billion-parameter model with 128K of context at 50 tokens per second. Against the M50’s published bandwidth, that is reachable only if the model is a mixture-of-experts design activating around four billion parameters per token. A dense model of the same size lands near 2 tokens per second, roughly one word every three seconds and useless for agentic coding.

The model library supports the inference. Qwen3.6-35B-A3B is a sparse design with three billion active parameters, listed alongside Qwen3.8-Flash-Next and a 27-billion-parameter model. Sparse designs are why a 40-billion-parameter model such as Recursive’s Flint can top local coding benchmarks. Lenovo is not shipping its heaviest weights as the default experience. It is shipping the ones that will feel fast.

The practical expectation follows. A 35B sparse model will feel genuinely quick. A dense 27B model will feel like 6 to 8 tokens per second, readable but sluggish inside an agent loop that calls tools repeatedly. The 125B model will load and answer, but slowly, which belongs to a wider shift toward Chinese models with very large context windows. This is a machine for running big models, not for running them briskly.

The Software Is the Actual Product

What Lenovo is really shipping is a token manager. The P7 is pitched as a personal token hub: one configuration screen where a user enters an API address and points coding tools, agent apps, and chat clients at the models running locally. Routine calls then stay on the device instead of drawing on a paid cloud quota, and the saving compounds where agent frameworks burn most of their tokens through retries and tool loops.

Beside it sits a model manager for viewing, starting, stopping, and switching installed models, and a dashboard showing hardware resources and task progress. A task mode keeps agent jobs running around the clock, with progress visible from a phone, tablet, or PC, which is what the 30-watt ceiling and the sub-35 dBA noise figure are for: the box is designed to be left on.

Lenovo’s Tianxi AI comes preinstalled, and an Auto mode routes each request between local and cloud models by complexity, so a trivial request never leaves the desk and a difficult one can escalate. The May documentation described two modes: an agent mode with Tianxi Claw preloaded and a large-model mode that opens the device to any external application as a local token supply.

That software layer is the difference between this and a mini PC. Whether the abstraction holds, particularly when a coding tool expects an exact API dialect, is the question the early-access program exists to answer. For a sense of how much there is to replace, Intelligent Living’s LLM API pricing comparison tracks what the same calls cost on the cloud.

How It Stacks Up Against the Alternatives

On capability, the P7 is not competing with the top of the local AI market. It competes on the ratio of capability to power and weight, where it has no direct rival.

Every machine in that table is a multi-kilogram box with a fan and a mains cable. The P7 is 300 grams, draws at most 30 watts, and Lenovo says it runs from an ordinary power bank at under 35 dBA. Those numbers describe an appliance, one that can be left running in a room where people work.

The closest philosophical rival is a cheaper mini PC with less memory. Intelligent Living has covered how Intel-based AI mini PCs reach 96GB without a five-figure price, and for many buyers those remain the rational choice. The P7’s argument is not that it wins on price or speed. It is that it fits in a bag, runs on 30 watts, and arrives configured.

Why Lenovo Chose Domestic Silicon

Both chips are Chinese, and one was custom-designed for the product. A device with no NVIDIA component has no allocation problem, no export-control exposure, and no dependency on a supplier constrained by data center demand. It fits a broader pattern of Chinese silicon built for inference, from Alibaba’s RISC-V model host to the accelerators filling Chinese server racks.

Both suppliers gain enormously. Houmo’s business rests on the thesis that inference belongs at the edge, and a Lenovo flagship is the most visible proof point available. CIX has bet close to a billion yuan on the agent CPU, a wager analysts have described as substantial, and Lenovo’s venture arm has published its own account of the CIX collaboration.

The rest of the family shows the intended scale. The AI Host mini, about half a Mac mini, offers 45 TOPS, 15 watts average, noise below 30 decibels, isolated data per user, and a first-release price of 2,999 yuan, roughly $420. That price is the best available signal for where the P7 lands, and it suggests Lenovo wants volume rather than margin.

What Remains Unproven

Lenovo says its in-house architecture, using mixed INT4 and INT8 quantization, averages 98.3% of a full-precision BF16 baseline across four benchmarks. INT8 typically loses very little accuracy; mixed INT4 losing only 1.7% is plausible but aggressive, and Lenovo has named neither the benchmarks nor the individual scores. The other claims deserve the same scrutiny:

• 50 tokens per second at 122B: consistent with a sparse mixture-of-experts model, inconsistent with a dense one by a factor of twenty. Lenovo has not specified which.
• 256K context at 125B: whether that fits inside 80 gigabytes of split memory once the key-value cache is allocated is unaddressed.
• 80 gigabytes of memory: the split between the two pools and its effect on sustained throughput, is undisclosed.
• 98.3% accuracy: no benchmark names, no per-test results, no methodology published.
• 30W maximum and under 35 dBA: plausible for a fanless 300-gram chassis, and the most easily verified claims on the sheet.
• The $420 comparison point: the AI Host mini price covers 45 TOPS and 15 watts, not the P7’s accelerator or memory load.

The absence of a price is itself informative. A company confident in its margins announces one at the preview stage. A company still negotiating memory costs, in a market where DRAM pricing has moved sharply against device makers, waits.

The Bottom Line

The Lenovo AI Host P7 is not a faster mini PC. It is a bet that the unit of AI consumption becomes a device rather than a subscription and that a 300-gram box drawing 30 watts covers the routine majority of what developers and office workers ask a model to do.

Judged on the published specification, the bet is coherent but unproven in three places. The two-pool memory architecture throttles the workloads the marketing leads with, and no measured throughput for a spanning model has been released. The 50 tokens per second figure depends on a sparse model shape Lenovo declines to name. And the price is unknown in a memory market that has been unkind to everyone building local AI hardware this year.

What is genuinely new is the software framing. Treating local inference as a quota replacement, with a manager that points existing tools at it by API address, is a more useful contribution than another TOPS figure. If that abstraction survives the early-access program, the P7 could matter beyond the Chinese market, and it would do so because of the token manager rather than the parameter count. It also feeds the home AI gold rush, now pulling serious hardware into spare rooms and wine cellars.

Frequently Asked Questions

How Much Will the Lenovo AI Host P7 Cost?

Lenovo has not announced a price. The only reference point in the family is the AI Host mini at 2,999 yuan, roughly $420, for 45 TOPS and 15 watts. The P7 adds a discrete accelerator, a larger memory configuration, and a bigger thermal envelope, so expect considerably more.

Can the Lenovo AI Host P7 Really Run 125B Models?

According to Lenovo, yes, locally, with a 256K context window and no cloud connection. The 125B figure assumes 4-bit quantization, so some quality loss is unavoidable, and a model that large must span both memory pools, which means it will generate text slowly.

Does the Lenovo AI Host P7 Need the Cloud?

Not for the functions Lenovo emphasizes. Coding assistance, document processing, speech recognition, and agent tasks run on the device. Auto mode deliberately escalates harder requests to cloud models, so the cloud is optional rather than absent.

When Does the Lenovo AI Host P7 Go on Sale?

Lenovo says only that it will go on sale soon and has given no date. The October 24 announcement opened an early-access program rather than a pre-order window. The May plan for crowdfunding on July 1 did not become general availability.