Divinci Local Inference

In-browser Gemma 4 inference via WebGPU for chat.divinci.app — model loads once, stays cached, shared across tabs.

Divinci Local Inference runs open-weight AI models directly in your browser, on your own GPU via WebGPU. Nothing you type goes to a server to be answered — the model is on your machine. Use it two ways: open the side panel on any page for an on-device assistant, or pick the local model in Divinci AI's chat at chat.divinci.app and have it served from your own hardware instead of the cloud. MODELS Choose the size that fits your machine. The download happens once, then the model is cached: • Gemma 4 E2B — ~2.9 GB • Gemma 4 E2B QAT — ~3.2 GB, best 4-bit quality • Llama 3.2 1B — ~0.9 GB • Qwen2.5 0.5B — ~0.5 GB • SmolLM2 360M — ~0.3 GB, smallest and fastest HOW IT WORKS The model is hosted in an offscreen document, so it loads ONCE per browser profile and stays warm across every tab — no reload when you switch pages. Model versions are pinned to a specific Hugging Face revision and never auto-update underneath you. WHY USE IT • Your conversations with the local model stay on your device • No per-token API cost — inference runs on hardware you already own • Works offline once a model is cached • Loads once, responds immediately in every tab thereafter OPTIONAL: SIGN IN Signing in to a Divinci account is entirely optional and off by default. It adds page-aware answers, grounded in Divinci's public-web index, and chat saved to your account. Both can be turned off individually in Advanced settings → Privacy. WHAT IT DOES NOT DO • Does NOT show ads, or use your data for advertising or cross-site tracking • Does NOT sell or rent your data • Does NOT send the CONTENT of the pages you visit — only a trimmed address and a one-way hash, and only while signed in with the panel open • Does NOT send anything about your browsing when signed out, or with the panel closed • Does NOT collect third-party analytics or telemetry PERMISSIONS • offscreen — hosts the model; MV3 service workers cannot use WebGPU directly • storage / unlimitedStorage — settings and conversation history, kept on your device; removes the quota cap on long histories • sidePanel — the docked assistant • identity — optional OAuth sign-in to your own Divinci account • host access to api.divinci.app and divinci-prod.us.auth0.com — used only for the optional signed-in features and the sign-in exchange • content script on all sites — draws the assistant on whatever page you open it on, and (only while the panel is open AND you are signed in) reads the page locally to compute a hash for the index lookup. Page content is not transmitted • externally_connectable — exactly one origin, https://chat.divinci.app, so our own web app can use the local model over a runtime port REQUIREMENTS • Chrome 113+ or another Chromium browser with WebGPU and offscreen support • A GPU supporting shader-f16 • Disk and memory sized to the model you pick — from ~0.3 GB for SmolLM2 up to ~3.2 GB for Gemma 4 E2B QAT • A one-time model download on first use OPEN SOURCE Apache-2.0. Source: https://github.com/Divinci-AI/gemma-gem Forked from kessler/gemma-gem, with attribution preserved in LICENSE. PRIVACY Local by default — chats with the on-device model never leave your computer. Signed-in features send only what they need: your basic profile at sign-in; a trimmed page address plus a one-way content hash while the panel is open; and your message when you ask for a page-aware answer. Sensitive sites — banking, webmail, healthcare, sign-in pages — are skipped. Full policy: https://divinci.ai/local-inference-privacy/