Running a 2.8-trillion-parameter model without a graphics card is possible
TL;DR
A developer just ran Kimi K3, a 2.8-trillion-parameter model, on an ordinary CPU with 8 GB of RAM. The project is called kimi-k3-in-c: 176 KB of code, no GPU, no framework. The principle: the model never sits in memory as a whole, it lives on disk and only loads the fraction it needs for each word it generates. You can reproduce the experiment at home, provided you have 1.7 TB of NVMe SSD and accept 10 to 32 seconds per generated word.
The wall everyone hits
The strongest AI models weigh hundreds of billions, sometimes trillions of parameters. Kimi K3, released by the Chinese lab Moonshot AI in late July 2026, holds 2.8 trillion. Getting answers comparable to GPT or Claude used to require a cluster of professional GPUs, out of reach for an individual or a small outfit.
The reason is mechanical: a model that size far exceeds the RAM of a consumer machine. The labs that train and serve it run on machines with several hundred gigabytes of VRAM, spread across multiple cards.
What one developer just demonstrated
Fareed Khan published kimi-k3-in-c, an inference engine for Kimi K3 written in pure C99. No PyTorch, no CUDA, no external math library. The repository holds up to inspection: Apache 2.0 license, active continuous integration, and a test suite that compares every computation against a PyTorch reference.
On the most modest machine tested, with 8 GB of RAM, the engine generates coherent text:
$ ./bin/k3 ~/k3model --preset laptop --prompt "The capital of France is" --gen 8
--- generated text ---
Paris.", "The Eiffel
The model answers correctly. It runs on a CPU. It fits in a laptop's RAM.
The principle, in one sentence
The model file weighs 1.56 TB on disk: no consumer RAM can hold it whole. kimi-k3-in-c works around the problem by keeping in memory only what it needs right now, and fetching the rest from disk at the precise moment it needs it.
Kimi K3 is a Mixture-of-Experts model: out of 896 internal experts, only 16 fire to produce a given word. kimi-k3-in-c exploits that property. It keeps a small part of the model resident in memory, and loads only the 16 required experts from disk for each word, rather than all 896 at once.
What it changes in practice
You can reproduce this on hardware you may already own: an AVX2-capable CPU, 8 GB of RAM, and above all 1.7 TB of NVMe SSD, the real prerequisite behind this project. Downloading the model alone takes several hours.
This project does not prove that a CPU replaces a GPU. It proves that a model designed for hundreds of gigabytes of VRAM can, with the right software architecture, run correctly on resources a thousand times smaller. The barrier is no longer only a matter of budget: part of it has become a matter of engineering.
One question remains, and we will dig into it separately: how fast. The answer tempers the enthusiasm of the headline considerably. How kimi-k3-in-c runs 2.8 trillion parameters on 8 GB of RAM explains the mechanism in detail, and Why a GPU is still essential explains why this is not usable day to day yet.