Gemma 4 Multimodal Agent v1.2.0

Secure, local-first RAG + Vision + Audio via WebGPU

Initialize AI Engine

Choose your local model execution backend

Embedding (MiniLM-L6)
Gemma 4 Multimodal

4B requires ~4GB. 12B is experimental and may exceed browser VRAM allocations.

Run models locally using Ollama with full hardware acceleration and MTP.

💡 Gemma 4 Speculative Decoding (MTP)

To run Gemma 4 12B faster with speculative decoding, create a model with the draft_num_predict parameter. Create a local Modelfile:

FROM gemma4:12b PARAMETER draft_num_predict 4

Compile it in your terminal:

ollama create gemma4-mtp -f ./Modelfile

CORS / Hugging Face Note: If connection fails:
1. Restart Ollama with origins allowed: $env:OLLAMA_ORIGINS="*" ; ollama serve
2. On Hugging Face, open the Direct Space URL (three dots → Embed this Space → Direct URL) and allow Insecure content in browser site settings to bypass Mixed Content blocking.