Developer Fareed Khan (FareedKhan-dev) published on GitHub the kimi-k3-in-c engine, written in portable C99. It runs the Kimi K3 model with 2.78 trillion parameters on a CPU with 8.24 GB of RAM, without a GPU or frameworks. The repository is public.
What is kimi-k3-in-c?
It is a minimalist inference engine whose complete code weighs only 176 KB. It compiles into a single binary from six C files with minimal dependencies.
Its goal is to demonstrate that a massive language model can run on consumer hardware, challenging the need for GPU clusters.
The trick: streaming from disk
The key is Kimi K3’s Mixture-of-Experts (MoE) architecture. The model has 93 layers and 896 routed experts, activating only 16 per token: about 104 billion active parameters out of the 2.78 trillion total.
The engine does not load the 1.56 TB of weights into memory. It keeps a dense part in RAM and streams the active experts from the NVMe SSD at each generation step, lowering memory usage to 8.24 GB.
The critical requirement: disk and speed
The most critical requirement is not RAM, but storage. A fast NVMe SSD with at least 1.56 TB free is needed. A traditional hard drive would make execution practically impossible, since about 135 GB per token must be read from disk.
With 8 GB of RAM, generation is about 32.7 seconds per token. A short response can take several minutes, making interactive use very difficult. Speed is the main practical obstacle.
Performance by RAM
Performance improves sharply with more memory. With 32 GB of RAM, the time per token drops to 24.2 seconds. With 64 GB, it falls to 19.8 seconds.
With 128 GB or more, the time per token falls to 5.6 seconds and the model fits entirely in memory. In all cases, the output is byte-identical; only the speed changes.
Availability and platforms
The engine is open source and available on GitHub. The reference platform is Linux x86-64; it also builds on macOS/arm64.
Native Windows support arrived in August 2026 via MinGW-w64/MSYS2, without requiring WSL. Before running it, about 1.56 TB of weights must be downloaded from Hugging Face.
An experiment, not a daily tool
The project is a technical demonstration of software optimization. The author presents it as a proof of concept: it works, but is impractical for daily use.
The value lies in showing that, with enough engineering, even the largest models can run on modest hardware. It is not a daily alternative, but an experiment for enthusiasts.
