'/%3e%3cpath%20d='M20%2040%20L28%2024%20L34%2034%20L39%2027%20L45%2040%20Z'%20fill='none'%20stroke='%23fff'%20stroke-width='3.2'%20stroke-linejoin='round'%20stroke-linecap='round'/%3e%3ccircle%20cx='32'%20cy='45.5'%20r='2.4'%20fill='%23fff'/%3e%3c/svg%3e)
1 Load the model (one-time download, then cached)
No models cached
First load downloads ~1.7 GB (Q4 + FP16 vision). Chrome / Edge 113+ with WebGPU recommended. Files are cached in your browser via the Cache API.
2 Image
Drop, paste or click to choose an image
PNG / JPG screenshots, scans, photos of documents
3 Prompt & run
Sampling defaults follow the model card: temperature 0.1, top-k 50, repetition penalty 1.1. Thinking off = faster direct transcription.