Well, yes, but actually no.LFM2.5 - 230M is truly impressive for it's footprint, and the technical background to deliver and run it embedded in a browser is trivial. The model picked up on training patterns in ~20M tokens, and easily manages structured tool calls.
But...
230M parameters for diverse domains is not enough. During training, the imprinting of different, generalized uses-cases compete for attention. The fidelity of interaction is also lacking, and for multi-turn conversations it tends to follow the trainined structure, instead of the actual context.
There are no cloud servers doing inference.
The model loads once and streams tokens right in your browser - WebGPU if available, otherwise CPU/WASM via wllama.
Besides the cool factor, running an on-device agent can be useful for:
Combining it with ARIA Roles and Semantic Elements can make navigating a complex UI a breeze.
This site is three demo storefronts sharing one on-device agent: a fine-tuned model that runs entirely in your browser and drives the interface with tools, rather than just answering in text.
Each demo gives it a different persona, catalog, and tool set - BrewCraft, Emporium, The Vendor - to show the same core loop adapts to very different UIs.
The live catalog and knowledge is injected as context, so replies stay grounded. Tools provide additional interactivity and context retrieval.
Voice input and voice replies are on-device models too (Whisper for speech-to-text, a small VITS model for text-to-speech) - nothing is sent anywhere for either.
Want more headroom? A 350M variant is available too, at ~1.5x the footprint - pick it from the model menu.
no accounts - no tracking