Decide Live runs multimodal decisions in-browser on EmbeddingGemma 2 and WebGPU with zero tokens
VIDRAFT and FINAL-Bench released Decide Live, a static web page that turns a camera frame, audio clip, image or text into a typed decision. The model, a fine-tune of google/embeddinggemma-2 (Apache-2.0), runs locally via transformers.js 4.3.1 and WebGPU with a WASM fallback, generating zero tokens and uploading nothing.
- Decisions come from softmax over embedding cosine similarity with scale 19.96, no decoding loop
- ONNX weights ship in 4-bit q4f16 and q4; fp16/fp32 exceeded the 1 GB Space limit
- Audio is processed in 16 kHz mono 2-second windows; clips below -55 dBFS are skipped
- An 18-image labeled test runs in-browser to report accuracy and compare prompt formats
Read next
AI