I’ve been running Gemma3 4b locally on ollama and it’s useful. I’m thinking about applications where a multimodal model could receive video or sensor feeds (like a security can, say).