WebGPU AI chat — how a browser tab runs a language model on your GPU
The short answer
WebGPU is the browser API that gives a web page direct access to your GPU's compute power, and it's what makes ShadowTalk AI possible: a quantized language model loads into GPU memory and generates tokens inside the tab — no server, no prompt leaving the device. It needs a modern Chromium browser (Chrome or Edge 113+), a one-time model download, and nothing else: no install, no drivers, no cloud GPU.
For thirty years the browser could show you things and, eventually, run games — but it couldn't think. WebGPU changed that, and this page explains what it actually does, what hardware and browsers support it today (November 2026), why the model weights come from ShadowTalk's own infrastructure, and where the honest limits sit.
What WebGPU actually does
WebGL let the browser paint; WebGPU lets it compute. It exposes modern GPU compute shaders to JavaScript, and language-model inference is exactly that — a long chain of matrix multiplications that GPUs do orders of magnitude faster than CPUs. ShadowTalk runs the model inside a background Web Worker talking to the GPU through WebGPU, so the tab stays responsive while tokens stream at hardware speed.
The model itself is quantized — compressed from full precision so that billion-parameter networks fit in consumer memory. That's why the on-device fleet is the 1B-3B class rather than a frontier model: a quantized 3B is genuinely useful and fits in an integrated GPU's reach; a frontier model does not fit in your tab, and honesty about that is part of the product.
Browser and hardware support, honestly
- Desktop: Chrome and Edge 113 or newer ship WebGPU by default — that's the supported path
- Android: current Chrome supports WebGPU, which is what makes fully-offline AI work on Android phones
- Safari and Firefox: incomplete or behind flags as of November 2026 — the PWA and cloud tiers still work there, but the on-device GPU path targets Chromium
- Hardware: integrated GPUs handle the 1B class comfortably; the practical ceiling in-browser is the ~3B class — low-end devices automatically fall back to cloud models instead of stuttering
- First run needs a connection to fetch the model (300MB-2.6GB, resumable shard-by-shard); every run after that is fully offline
Why the model weights come from ShadowTalk's own infrastructure
A surprising number of browser AI demos pull their model weights from third-party CDNs, which means your first load trusts a supply chain you can't see. ShadowTalk serves its own app shell and model weights from its own infrastructure with zero third-party requests — the offline hardening work exists so the whole stack, from fonts to shards, behaves like one self-hosted product. Downloads are sharded and resumable, so a dropped connection never restarts a 2GB file from byte zero.
What the GPU path changes for privacy
When inference happens on your GPU, the prompt never crosses the network — there is no server to see it, no log to hold it. That's the on-device tier. Sync is a separate, opt-in system: default chats are stored server-side when Private Mode is off, and when you switch Private Mode on, stored chats are end-to-end encrypted (AES-256-GCM) with keys derived on your device. The GPU path is private by physics; Private Mode is private by cryptography — you can have both.
Check your browser is WebGPU-ready and start
- 1Check your version: Chrome or Edge 113 or newer (chrome://version), on desktop or Android
- 2Open shadowtalk-ai.com/chatbot — no install step exists
- 3Pick the on-device model and let it download over Wi-Fi once; shards resume if the connection drops
- 4Chat — if your device is below the GPU bar, ShadowTalk automatically falls back to cloud models rather than stalling
Frequently asked questions
Keep reading
Free to use — no account, no card.
Works offline. Syncs when you're back. Your keys, your models, your chats.
Run WebGPU AI — free