A 2.78-Trillion Parameter Model Just Ran on 8 GB of RAM

You can now run a 2.78-trillion-parameter AI model on a Linux machine with 8 GB of RAM and zero GPUs.

Developer Fareed Khan released kimi-k3-in-c, a 176 KB engine written in plain C99 that runs Moonshot's Kimi K3 model. It streams the model's 1.56 TB checkpoint directly off disk, running on a single CPU without external frameworks or BLAS libraries.

Why it matters: It turns hardware requirements into a sliding dial instead of a brick wall. By keeping a small core in memory and streaming 1.45 TB of expert weights off storage, the tool delivers byte-identical outputs whether you budget 8 GB or 224 GB of RAM.

Generating a token takes roughly 32 seconds at 8.24 GB of RAM, dropping to 10 seconds at 128 GB. Because this uses Kimi K3's raw base model, prompts generate text continuations rather than formatted chat responses.

It won't replace a server cluster for speed, but it proves data-center-sized models don't strictly require data-center hardware.