The PyTorch team has released a backend for ExecuTorch that runs PyTorch models on a computer’s graphics chip inside a web browser, with no detour through another framework’s file format. A developer who wanted a model in a browser until now usually had to convert it to GGUF, ONNX, or TFLite first. The new route keeps the model inside PyTorch’s own tooling from export to delivery, according to a post on the PyTorch blog dated 10 November 2026.
The workflow is short. A developer exports with torch.export into a single ExecuTorch file that ends in .pte, and that file runs in any browser that supports WebGPU, the web standard for reaching the graphics chip. The same file also runs outside the browser, on Metal for macOS, Vulkan for Linux and Android, and D3D12 for Windows. Shipping it means putting the file on a CDN and loading it from a web page. The blog says nothing gets installed, no server is involved, and whatever the user types, shoots, or says never leaves the device.
The speed claim comes from a benchmark of Qwen3-0.6B and a quantized Llama 3.2 1B in Chrome on an Apple M1 Pro laptop chip. The PyTorch team compared the graphics path against ExecuTorch’s own processor path in the browser (WebAssembly with XNNPACK) and found WebGPU 4 to 7 times faster at prefill, the stage where the model reads the prompt. It was 2 to 10 times faster at decode, where the model writes its answer. Each figure is the median of five runs after three warmups, across contexts from 128 to 8K tokens.
Read the comparison for what it is. The baseline is the team’s own processor backend, not another browser runtime such as ONNX, and the post publishes no head-to-head against one. The two runs also differ in setup: WebGPU stored its weights in 4 bits and kept activations at full precision, while the processor path paired 4-bit weights and 8-bit activations. Every model in the speed test sits at around a billion parameters or below, and all of it ran on one machine.
The range of models is wider than the benchmark. The team lists Llama 3.2, Qwen2.5, Qwen3, and Gemma 3 for text, SmolVLM for questions about images, Whisper for speech-to-text, and Phonon-31M-EN for text-to-speech. Vision models include Depth Anything V2, SAM 2, D-FINE, Real-ESRGAN, and MODNet for background removal. LoRA and DiReFT fine-tuning also ran on Llama 3.2 1B and the two Qwen models, which would let a browser app adapt a small model to one user without sending anything to a server. A playground page runs many of these, and the Hugging Face page for each model carries the .pte file and the export recipe.
The business angle sits in who pays. Inference that runs in the visitor’s tab moves the compute bill from the company’s servers to the visitor’s laptop, and it removes the privacy question about where a prompt goes. The PyTorch team itself calls the backend early development, with faster and more stable builds, plus more operators, data types, and models, still to come.
A product team with a model near the billion-parameter mark can try this on an afternoon’s export, but the M1 Pro numbers say little about the older and cheaper machines its visitors will use. Test there first, and decide after.
Reported by the PyTorch blog on 10 November 2026.
