A small language model chatting locally in the browser: SmolLM2-360M, LoRA fine-tuned, runs over WebGPU through a patched Rust/WASM runtime with no inference server, while a second Rust/WASM module retrieves evidence from a fictional research-station knowledge graph. Every answer shows the facts it used. The technique first shipped inside stix-meta-explorer’s STIX assistant; this is it on its own, with the whole pipeline from graph generation through training, evaluation and browser proof published. Experimental, desktop only.
llm-in-browser
A small language model chatting locally in the browser: SmolLM2-360M, LoRA fine-tuned, runs over WebGPU through a patched Rust/WASM runtime with no inference server, while a second Rust/WASM module retrieves evidence from a fictional research-station knowledge graph. Every answer shows the facts it used. The technique first shipped inside stix-meta-explorer's STIX assistant; this is it on its own, with the whole pipeline from graph generation through training, evaluation and browser proof published. Experimental, desktop only.
Published Project 2026 – present
- Related
- stix-meta-explorer