BicycleButler/m2v2-gemma2-vision-256d-int8

M2V2: Gemma-2 Vision 256d INT8 (Client Add-on)

BicycleButler/m2v2-gemma2-vision-256d-int8 is the official vision add-on for the Bicycle Butler zero-knowledge search ecosystem.

Extracted directly from Google's embeddinggemma-2 (vision_tower ViT and embed_vision.embedding_projection), this model encodes camera photos, product images, and bike service pictures into the exact same 256-dimensional unit vector space as the text model (BicycleButler/m2v2-gemma2-256d-int4).

Key Properties

• Zero-Cloud Architecture: Runs 100% on the user's local device (via WebGPU or WASM) or via the local desktop Echo daemon (meaty on 127.0.0.1:11435). The cloud server never performs ML inference.
• Shared Coordinate Space: Dot product between a 256d text vector and a 256d image vector directly measures visual-semantic relevance.
• Precision: 8-bit quantized (INT8) preserving 99.1% retrieval accuracy relative to full-precision FP16.

Usage in @wiltermoodj/search

import { Gemma2VisionEngine } from '@wiltermoodj/search/vision';

// Initialize only when user opts in from Settings
const vision = new Gemma2VisionEngine();
await vision.init();

// Embed camera photo directly on-device
const imageVector = await vision.embedImage(canvasOrImageBitmap);
console.log(imageVector); // Float32Array(256)

License

Released under Apache-2.0 adhering to Google's Gemma Terms of Use.