On 4 November 2026, Pandaily reported that Lenovo Kaitian had launched a 1.6-litre AI desktop for China’s Xinchuang IT application innovation market. Pairing a Zhaoxin KX-7000 processor with a Houmo Liqing LQ50 neural processing unit, or NPU, the machine is designed to run models of around 35 billion parameters on the device. That puts the processing for a class of large language models inside the desktop rather than on a remote server.
Key points
- The standard configuration pairs an eight-core Zhaoxin CPU with a Houmo NPU rated at 160 TOPS of INT8 compute.
- Lenovo reports about 28.99 tokens per second for Qwen3.6-35B-A3B using W4A8 quantisation on the 24GB NPU.
- The host runs Kylin OS and includes Lenovo’s Tianxi AI Pro agent software.
- Lenovo reports case temperatures of about 41 degrees Celsius under full load in a 25-degree room.
The LQ50’s memory and 160 TOPS
The desktop uses an eight-core Zhaoxin KX-7000 alongside the Houmo LQ50. Its standard NPU configuration is rated at 160 TOPS of INT8 compute, with an option of up to 320 TOPS. TOPS counts trillions of arithmetic operations a chip can perform each second, rather than words a model can produce each second. The LQ50 also has 24GB of dedicated memory and 153.6GB/s of memory bandwidth, both relevant to keeping model data available as the chip generates text.
Lenovo said it tested the host with open-source models including the Qwen3.5 series and Qwen3.6-35B-A3B. For the latter, it reported text generation at about 28.99 tokens per second using the 24GB NPU and W4A8 quantisation. Quantisation stores numbers used by a model at lower precision, rather like using fewer digits for a measurement. That choice helps fit a large model into limited chip memory while making the reported speed specific to the tested setup.
According to Lenovo, the host can accept as many as 200,000 tokens of input for extended documents and conversations across multiple turns. Input length and generation speed describe different parts of the work: one concerns how much text the system can take in, while the other concerns the pace at which it writes a response. The reported generation rate comes from the Qwen3.6-35B-A3B run with W4A8 quantisation.
Summarising a long document could take place on the desktop if the text fitted within that supported input length. Drafting a response from the same material could likewise use local inference rather than cloud tokens, as Lenovo describes for the host. Those activities would depend on the model and the configuration used for the task.
Tianxi AI Pro runs on Kylin OS
The host runs Kylin OS and arrives with Lenovo’s Tianxi AI Pro agent software installed. Lenovo pitches it as an always-on digital employee, designed for scheduled work such as organising files, searching local knowledge, drafting documents and summarising information. In its description of the setup, inference takes place on the machine rather than through cloud tokens, keeping the data used for those tasks local.
At a Tianxi AI Pro event, testers gave the system an instruction to check a mobile carrier app. It operated a test phone, measured the time taken for an SMS verification code to arrive and reported the result. That demonstration concerned a task involving the software agent and a phone. The reported Qwen generation rate, by contrast, concerns text produced by a particular model running locally on the NPU.
Houmo supplies software for quantising and compiling models, managing their runtime and profiling their performance. Those tools give developers a way to prepare models for the accelerator and examine how they run on it. For companies building around the host, the combination of that software, Kylin OS and the two Chinese chips defines the environment in which local applications would run.
Cooling the 1.6-litre Kaitian host
The CPU and NPU have separate air channels behind a hexagonal front panel, with turbine fans moving air through the case. In a 25-degree room, the company puts the case temperature at about 33 degrees Celsius while idle and 41 degrees under full load. In a 35-degree room, it reports about 46 degrees under full load. It puts noise at 19.4dB when idle and about 29.5dB during full NPU inference.
The unit weighs about 1.5kg and has two independent 2.5GbE network ports. It is separate from the Lenovo AI Host P7, which Lenovo previewed in October with a Houmo M50 and a CIX P1 instead of the LQ50 and KX-7000 used in the Kaitian machine.
