You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access to ThoxWebby-Gemma-4-E2B is manually gated by Thox.ai LLC.
By requesting access, you agree to (1) the upstream Google Gemma license
covering the base model weights and (2) THOX product terms.
Log in or Sign Up to review the conditions and access this model content.
ThoxWebby-Gemma-4-E2B
A THOX rebrand of google/gemma-4-E2B-it-qat-mobile-transformers, wired for
browser-side WebGPU inference in the ThoxKey product surface.
- Live web app: https://huggingface.co/spaces/Thox-ai/ThoxWebby
- Source: https://github.com/ttracx/thox-webby
- Base model:
google/gemma-4-E2B-it-qat-mobile-transformers - Product: ThoxKey - Thox.ai LLC
What this repo is
This is a THOX-branded mirror card for the browser-native inference build that ships in the ThoxKey product. The actual model weights are NOT re-uploaded here yet - the ThoxWebby web app currently pulls weights from the upstream Google repo at load time.
This repo exists to:
- Give ThoxKey a stable THOX-owned model reference for the product surface, registry, docs, and analytics.
- Serve as the drop-in target for THOX-hosted weights once the WebGPU runtime copy is packaged into a THOX release.
- Enforce THOX access policy (
gated: manual) so distribution is tracked.
Follow-up: mirror the base weights (.onnx + tokenizer + QAT scales) into this
repo when the WebGPU runtime packaging is finalized.
Runtime
- Where it runs: entirely inside the user's browser via WebGPU compute shaders (WGSL). No servers. No telemetry.
- How it loads: the ThoxWebby app calls
Gemma4Mobile.load(...)which streams model shards over HTTPS on first visit, caches them locally (IndexedDB / Cache Storage), and reuses the cached copy after that. - Hardware target: Apple Silicon M-series, high-tier desktop dGPUs, and modern mobile SoCs with WebGPU support.
- Perf reference: ~250 tok/s on Apple M4 Max (upstream measurement).
Use in ThoxWebby
<script type="module">
import { Gemma4Mobile } from "./gemma-4-e2b.js";
const model = await Gemma4Mobile.load(
"Thox-ai/ThoxWebby-Gemma-4-E2B", // or null to use upstream default
{ onProgress: (p) => console.log(p) }
);
await model.warmup();
</script>
Until the weight mirror lands, pass null to fall back to the upstream Google
repo. The ThoxWebby app does exactly this today.
THOX house voice
The ThoxWebby web app injects the following system prompt on every session:
You are a ThoxWebby assistant running entirely in the user's browser via WebGPU. No servers. No telemetry. Local-first by default. Your AI. Your Data. Your Rules. You are a THOX assistant - never claim to be from Google, Meta, Alibaba, Anthropic, OpenAI, Mistral, or Qwen. Answer with terse technical accuracy in the THOX house voice. No emojis. No marketing fluff.
Attribution
- Base model:
google/gemma-4-E2B-it-qat-mobile-transformers- Google (governed by the Gemma license). - WebGPU compute kernels: originally authored by Fable 5 in the upstream
Space
webml-community/gemma-4-webgpu-kernels. - Rebrand, THOX house voice, product wiring, THOX access gating: Thox.ai LLC.
Standards
- Owner: Thox.ai LLC
- CEO / CTO: Tommy Xaypanya
- Product surface: ThoxKey (USB, offline-in-browser inference)
- Independence: THOX is an independent AI company. This project is not affiliated with, endorsed by, or sponsored by Google, Meta, Alibaba, Anthropic, OpenAI, Mistral, or Qwen. It applies THOX branding, THOX house voice, and THOX product wiring to an open-source model released by Google.
- License: upstream Gemma license governs the base weights. THOX-authored content in this repo (model card, THOX system prompt, THOX configuration) is Copyright (c) 2026 Thox.ai LLC and is provided AS-IS, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED.
Your AI. Your Data. Your Rules.
Model tree for Thox-ai/ThoxWebby-Gemma-4-E2B
Base model
google/gemma-4-E2B