Artifacts

Transcribes or translates speech from an audio file, URL or microphone with Whisper running inside the browser tab.

Estimates how much GPU memory a language model needs for a given size, quantisation and context length.

Streams live microphone audio through Whisper on the GPU in your browser and prints the words as you speak.

Transcribes a recording in the browser and labels which speaker said each segment.

Runs a Gemma 4 model in the browser with custom WebGPU kernels, so chat needs no server.

WebGPU-accelerated Whisper that turns microphone or file audio into text locally in the browser.

Speech recognition with the smaller Distil-Whisper model, executed locally in the browser.

Converts a document page image into structured Docling markup using a 258M model run in the browser.

Points your webcam at a scene and has a small vision-language model describe it continuously, locally on WebGPU.

Chat playground that downloads open models and runs them locally on WebGPU with WebLLM.

Runs Phi-1 and Phi-2 text generation in WebAssembly through Rust Candle with no server.

Runs the llama2.c small Llama model in WebAssembly with Rust Candle, generating text in the page.

Click points on an image and get object masks from Segment Anything running in the browser.

Multimodal chat with Qwen 3.5 vision models running locally via Transformers.js and WebGPU.

Runs Microsoft's Florence-2 vision model in the browser for captioning, detection and OCR tasks.

Demonstrates a small model calling tools from the browser using Transformers.js and WebGPU.

Translates text between many languages with an NLLB model executed inside the browser through Transformers.js.

A compact reasoning-capable Nemotron model that runs in the browser on WebGPU.

Open-weight text-to-speech that downloads the model once and runs in your browser.

Multilingual text-to-speech with voice cloning that runs entirely in the browser on CPU.

An on-device research agent built on LFM2.5 2.6B that runs through WebGPU in the browser.

Ask questions about a picture using the Moondream vision-language model running locally on WebGPU.

Loads the open-weight GPT-OSS 20B model into WebGPU memory so you can chat without a server.

Runs the very small LFM2.5 230M language model with custom WebGPU kernels in the tab.