AI tools that run in your browser — and never upload your files

There is a quiet split in the AI tool market. Most products send your file to a server and return a result: that architecture is why they have upload boxes, size limits and subscription tiers. A smaller group downloads the model to your device and runs it there, which removes all three.

The list below is the second kind. The test I use is simple: load the page, disconnect from the network, and try the feature. If it still works, the model is genuinely local.

Editing and creative work

Why this became possible

  1. WebGPU exposed the GPU to the browser without a plugin, so vision models stopped being a CPU-bound compromise.
  2. Quantization brought model size down by three to four times with little visible loss on editing tasks.
  3. WASM SIMD and threads made the fallback path acceptable on machines without a usable GPU.

How to tell them apart

Notes on the engineering side — cold starts, bundle budgets, device variance — are collected at Running vision models in the browser.