Run AI models in your browser — free, unlimited, private.
Live: https://yeeeeezus.github.io/ai-lab/
A minimal inference playground built on transformers.js (Hugging Face, Apache-2.0). Model weights download once from the Hugging Face Hub — then inference runs on your own device via WebGPU when available, WASM otherwise:
- Text generation — SmolLM2-135M-Instruct (Apache-2.0, ~90 MB q8) and Qwen2.5-0.5B-Instruct (upstream Apache-2.0, ~350 MB q4). Real chat-template inference with token streaming, temperature and max-token controls, and an interruptible Stop button.
- Sentiment analysis — DistilBERT SST-2, ~30 MB quantized
- Named entity recognition — BERT CoNLL-03, people / organizations / locations highlighted in your text
No API key. No server. No rate limits. Nothing about your input leaves the page — model files are fetched from the Hub, predictions are computed locally.
Hosted model APIs meter you: per-request billing, daily caps, or retirement — GitHub's free Models inference API was wound down in 2026. Small models running client-side are the durable alternative for demos, tools and privacy-sensitive inputs. ai-lab is the smallest honest version of that idea: one static page, proven models, device-appropriate backend. The generative models are chosen for what a browser can actually run well — a 135M instruct model is genuinely useful for short drafts and lists, and a 0.5B trades download size for a bit more coherence.
Nothing to build — open index.html, or:
$ python3 -m http.server 8000First model load needs network access to huggingface.co; after that the browser cache takes over.
MIT — models remain under their authors' licenses (see links above).