Homelabber Runs Massive AI Model on CPU Using Optane Memory and Blower Fans

A computer hobbyist has successfully run Moonshot AI’s massive 2.8-trillion-parameter Kimi K3 AI model at home using only a CPU and discontinued second-generation Intel Optane memory sticks. The project achieved this heavy workload without using a graphics card, but the builder warns that the extreme heat generated by the Optane modules can burn out nearby memory.

Running a model of this size at home is very slow, generating only about 0.85 tokens per second even with heavy 1-bit quantization. To keep temperatures under control and prevent the hardware from melting, the builder had to use dual 12-watt blower fans aimed directly at the memory modules. The builder advises that active cooling for these memory slots is mandatory.

The Kimi K3 model is designed for commercial data centers and normally requires specialized infrastructure. Because regular memory prices have skyrocketed, buying secondhand Optane memory sticks is one of the few affordable ways for a home builder to reach the massive 610GB of RAM required by the smallest community version of the model.

Intel launched its Optane Persistent Memory 200 Series in 2020 in 128GB, 256GB, and 512GB capacities to work with its 3rd-generation Xeon Scalable processors. While Intel officially discontinued its Optane business in mid-2022, these memory sticks provide a cost-effective way to expand capacity by running in App Direct Mode, which gives software direct control over where the model's data lives.

Performance drops further when using a higher 3-bit quantization, lowering speeds to 0.75 tokens per second. At that speed, a 50,000-token task takes roughly 13 hours to complete.

The severe heat issue matches Intel's official technical specifications, which state that the modules require data-center-level forced airflow to stay cool. Each Optane module consumes 12W to 15W of power, meaning a four-stick setup places up to 60W of heat directly next to the system's regular memory. The Optane modules begin to throttle performance at 83 degrees Celsius and trigger a thermal shutdown at 85 degrees Celsius.

The builder plans to add graphics cards and fine-tune thread settings in the future to improve performance.

Summary

Running a massive, data-center-grade AI model at home is possible using discontinued server memory and a standard CPU, but hobbyists must manage extreme hardware heat to avoid destroying their systems.