joeysalmons/slm-in-a-box

slm-in-a-box is an experiment that packs a small AI language model into a single HTML file, so anyone can chat with it just by opening that file in a web browser like Chrome. There is no installation, no server to run, and no internet connection needed once the file is downloaded. The model at the center of this project has 50 million parameters, which is tiny compared to the large models behind tools like ChatGPT, but it is small enough to fit inside a self-contained web page along with everything needed to run it. Because the whole thing lives in one file, it works completely offline after the initial download, and it can even be carried around and run from a USB stick on a different computer. The HTML file includes a basic chat interface with several adjustable settings: temperature, top-p, top-k, maximum tokens, and repetition penalty, which are all common controls for tuning how creative or repetitive the model's replies are. It also has a system prompt field, an editable conversation history, and buttons to stop, regenerate, or copy responses. On a modern laptop running Chrome, the author reports speeds of around 50 tokens per second. There is an experimental option to use WebGPU instead of the default CPU-based backend, which runs through WebAssembly, but the README warns this will likely freeze the browser tab, so it recommends sticking with the CPU option. For people who want to build the file themselves rather than download the prebuilt one, the repository includes a Python script, build_supra_chat-v2.py. Running it requires first downloading a specific set of model files, an ONNX version of the Supra 1.5 50M Instruct model along with its tokenizer and config files, from Hugging Face, and placing them in the same folder as the script.