llama2.c

A small C program that runs small Llama 2 language models, with training done in PyTorch.

Stars: 20.2k
Forks: 2.7k
Licence: MIT
Language: C
Last update: 2.3 years ago
Type: Other

No commits in over a year — check for forks or alternatives before adopting.

What it is

This project runs Llama 2 style language models using a single short C file with no dependencies. The README says you train a model in PyTorch and then run it with that C program. It is mainly built for minimalism and learning how inference works, not for running large models quickly.

Download the 15M parameter TinyStories checkpoint and generate short stories on your own machine
Sample text with custom temperature, top-p, and prompt settings using command line flags
Export Meta's Llama 2 checkpoints to the llama2.c format and run the 7B base model, as described in the README
Chat with a Llama 2 chat model using the -m chat flag

Great for

Developers who want to see how LLM inference works in plain C
Students learning the Llama 2 architecture end to end
People experimenting with tiny models trained on narrow datasets like TinyStories
Anyone who wants a minimal inference engine with no external dependencies

Not ideal for

People who need to run 7B or larger models, since the README says the code only does float32 inference and 13B+ does not currently work
People who need fast production inference, since the README describes it as educational and float32-only
People who want a polished, supported product, since the README says the project is young and moving quickly

Step-by-step guide: how to use llama2.cInstall, first run, examples, tips and FAQs

llama.cpp vs llama2.c llama2.c vs llama3 from scratch

How to get it

1
Clone and build from source
You'll need: C compiler with make available · Python with pip for exporting Meta checkpoints (requirements.txt)
Clone the repository
git clone https://github.com/karpathy/llama2.c.git
Open the repository folder
cd llama2.c
Download the 15M parameter TinyStories model
wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories15M.bin
Compile and run the C code
make run
2
Run it
- Clone the repository and change into its folder
- Download the stories15M.bin checkpoint (about 60MB)
- Compile with make run, then run ./run stories15M.bin to see a streamed sample
- Try the larger stories42M.bin model for more coherent output
3
Get the most out of it
- Use the Makefile compile flags, which the README says can significantly speed up inference
- Tune sampling with -t for temperature or -p for top-p, but not both at once, as the README recommends
- Use the -i flag to give the model a prompt and -n to set the number of steps
- Export a Meta Llama 2 chat model with export.py and run it with -m chat

Downloads

Source code (.zip)master branch · straight from GitHub

Strengths

Inference is a single short C file with no dependencies, which makes it easy to read
The README includes a full train-plus-inference workflow in one repo
Small models trained on narrow data can produce surprisingly coherent output, as shown in the examples

Trade-offs

The README says only fp32 inference is supported, so large models need a lot of memory and run slowly
13B and larger models do not currently work according to the README
Code Llama support is incomplete and the README calls it buggy

How it compares

Project	Stars	Type	Level	Updated	Best for
llama2.c (this)	20.2k	Other	intermediate	2.3 years ago	Developers who want to see how LLM inference works in plain C
llama.cpp	133k	Library	intermediate	4 days ago	Developers who want to run open-weight models locally without a cloud API

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.