Running an LLM agent entirely in your browser

Short summary

A developer fine-tuned LiquidAI's LFM2.5 (230M and 350M parameters) into a browser-only front-end agent that handles tool selection, argument binding, reference resolution, and RAG without any server or API key. The model learns interaction patterns rather than domain facts, so the same weights power multiple storefronts with zero retraining. Key design choices include a frozen tool roster, RAG-as-a-tool architecture, and grammar-constrained decoding, with synthetic data distilled from Qwen3 30B.

• •230M/350M parameter model runs entirely in-browser with no server or API costs
• •Trained on interaction patterns not domain facts—one model powers multiple storefronts
• •Uses grammar-constrained decoding and RAG-as-a-tool for reliable tool calling

Generated with AI, which can make mistakes.