colibri
Tiny C engine that runs huge MoE models on the machine you already own — GLM, Kimi K3, DeepSeek, Qwen, Apache-2.0
Kimi K3 in C is an open-source local llm project by FareedKhan-dev (FareedKhan-dev/kimi-k3-in-c on GitHub). It has 12,114 GitHub stars and 1,873 forks, is released under the Apache-2.0 license, and was last updated on GitHub on 2 October 2026.
Kimi K3, 2.78 trillion parameters, running on one CPU in 8 GB of RAM — portable C99, no GPU
Seven C files, a compiler and OpenMP — the build takes seconds.
Run the doctor script to size RAM to a preset and measure storage.
Run the test suite first — it needs nothing but the repository.
Only then commit to the 1.56 TB download, pack the trunk, and run.
The full Kimi K3 model on one CPU, streaming weights from disk.
No GPU, no BLAS, no ML framework — Linux, macOS/arm64 and Windows all build.
Checks the toolchain, sizes RAM to a preset, measures storage, prints the next command.
The test suite proves the engine matches its reference with no checkpoint at all.
Presets from laptop (8 GB) to server pin more trunk layers as RAM grows.
Ids in, ids out, no tokenizer, JSON results — plus cache traces to replay offline.
Download the source and move into the folder.
git clone https://github.com/FareedKhan-dev/kimi-k3-in-c.git
cd kimi-k3-in-cCompile the engine — seconds, seven C files.
make -jRun the test suite before downloading anything.
make testClose neighbours on the Local LLM shelf: colibri — Tiny C engine that runs huge MoE models on the, and llama.cpp — LLM inference in C/C++. The full shelf is the Local LLM tools in the catalogue. Before running a model locally, check how much VRAM you need for local LLMs — model size has to match your hardware.
A portable C99 engine that runs Moonshot AI’s 2.78-trillion-parameter Kimi K3 on one CPU in as little as 8 GB of RAM.
About 26.5 seconds per token at 8 GB, about 5.6 seconds per token at 128 GB or more — a proof of fit, not a chat setup.
A CPU with AVX2 + FMA (or NEON on arm64), 8 GB of RAM, and ~1.7 TB free for the 1.56 TB checkpoint.
No — no GPU, no BLAS and no ML framework. That is the whole point.
Yes, Apache-2.0 licensed. The Kimi K3 checkpoint itself is Moonshot AI’s open-weight release.
Tiny C engine that runs huge MoE models on the machine you already own — GLM, Kimi K3, DeepSeek, Qwen, Apache-2.0
LLM inference in C/C++
GPT4All — run local LLMs on any device: a free desktop chat client and Python SDK that run open models entirely on your CPU, no GPU or cloud required
Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Self-evolving general AI agent running fully locally, a Manus-style alternative.
Cloud-native runtime for building and running agentic AI in production.
The most-starred other Local LLM projects in the LibreAI catalogue — the realistic shortlist if Kimi K3 in C is not the right fit.
The simplest way to run open language models on your own machine.
User-friendly AI Interface (Supports Ollama, OpenAI API, ...)
LLM inference in C/C++
A high-throughput and memory-efficient inference and serving engine for LLMs
🤯 LobeHub is your Chief Agent Operator, organizing your agents into 7×24 operations by hiring, scheduling, and reporting on your entire AI team.
Nearby Local LLM projects in the catalogue, closest in stars to Kimi K3 in C — keep exploring the shelf either side of it.
Typescript/React Library for AI Chat 💬🚀
KoboldCpp - Run GGUF AI models locally with a KoboldAI UI. One File. Zero Install
Access large language models from the command-line
Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Open bilingual (Chinese-English) 6B chat LLM with 32K-context variant and low-memory INT4 inference.
vLLM-Omni — efficient inference for omni-modality models: image, audio, video, diffusion and policies.
Free and open — grab the source, star the repo, and start creating.