ExecuTorch brings GPU-accelerated PyTorch models to the browser

PyTorch models can now run in a web browser using the GPU, without first being converted to another framework’s format. The new ExecuTorch WebGPU backend is aimed at making local AI models easier to run on the web, and PyTorch says it is 4 to 10 times faster than ExecuTorch’s previous CPU route.

Watch Desk analysis

What happened

The backend runs exported .pte model files using WebAssembly and WGSL compute shaders. PyTorch says it supports a range of AI workloads, including large language models, vision-language systems, depth estimation and speech-to-text.

That means developers can use GPU acceleration for these models inside a browser while keeping them in the ExecuTorch and PyTorch toolchain. The claimed speed-up is measured against ExecuTorch’s previous CPU path, not against every browser or native implementation.

Why it matters

Running models in the browser can make AI features easier to try and deploy without asking users to install a separate application or send every task to a remote service. GPU acceleration may also make browser-based models more practical for interactive work. The useful change here is not just that a model fits inside a tab; it is that developers can bring existing PyTorch models along without converting them to a different framework format.

Our read

This is a promising step towards local AI that feels like ordinary web software. The reported 4-to-10-times improvement is a striking starting point, though it compares the new backend with ExecuTorch’s CPU route. Real-world speed will depend on the model, browser and device. Still, getting a broad set of model types onto a browser GPU without a framework detour is an appealing bit of engineering, not merely a novelty demo.

What to watch

• Which browsers and devices can run the backend reliably.
• How performance varies across supported model types and hardware.
• Whether developers build useful browser-based features with the new backend.

Discussion spark: Would you trust a browser to run useful AI models locally, or do you still prefer the flexibility of a cloud service?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.