Local WebGPU LLM Benchmark
Execute quantized open-weights LLMs entirely in your browser via WebGPU to measure pure decode throughput and latency.
Probing WebGPU Adapter
Querying browser GPU capabilities and hardware limits...
About In-Browser WebGPU LLM Inference
Your prompts, inputs, and generated tokens never touch an external inference server. The model weights execute directly within your browser process via hardware-accelerated shaders.
On first launch, model weights download from HuggingFace CDN into browser Cache Storage. Subsequent launches reload weights locally with zero repeated network downloads.
Benchmark real decode throughput (tokens/second) and time-to-first-token (TTFT) on standardized prompt suites to evaluate whether on-device AI fits your user workloads.
Browser & Hardware Support Matrix
WebGPU requires a modern desktop browser and compatible graphics drivers (Vulkan, Metal, or Direct3D 12).
| Platform / Browser | Minimum Version | Status | Configuration Notes |
|---|---|---|---|
| Google Chrome | 113+ (Desktop) | Supported | Enabled by default on Windows, macOS, Linux, and ChromeOS. |
| Microsoft Edge | 113+ (Desktop) | Supported | Enabled by default. Inherits Chromium WebGPU backend. |
| Apple Safari | 26+ (macOS / iPadOS) | Supported | Native Metal backend. Requires modern macOS/iPadOS releases. |
| Mozilla Firefox | 115+ (Windows) | Experimental | Requires enabling dom.webgpu.enabled in about:config. |
| Mobile Browsers | iOS Safari / Android Chrome | Unsupported | OS memory limits and shader sandbox caps prevent LLM execution. |
Quantized Model Catalog
Specifications and resource footprints for models supported in the WebGPU playground.
| Model | Parameters | Quantization | Download Footprint | Declared VRAM | Runtimes | Recommended Use |
|---|---|---|---|---|---|---|
SmolLM2 360M InstructRecommended smollm2-360m | 360M | q4f16_1 | ~280 MB | ~376 MB | WebLLMTransformers.js | Fastest download. Start here. |
Qwen2.5 Coder 1.5B qwen2.5-coder-1.5b | 1.5B | q4f16_1 | ~1100 MB | ~1630 MB | WebLLM | Best code quality in this range. |
Llama 3.2 1B Instruct llama-3.2-1b | 1B | q4f16_1 | ~880 MB | ~880 MB | WebLLM | General reasoning & chat |
