Frigate user hunts vision model

This is the right framing for local AI at the edge: the best model is the one that gives useful answers fast enough to keep the automation loop intact.

• –Frigate is built around local AI object detection and Home Assistant integration, so latency and efficiency matter more than raw model size
• –Gemma 4 E2B is already a sensible edge-sized baseline, and the reported slowness suggests the A380 is the bottleneck
• –For tasks like counting lights, a smaller or more tightly scoped vision model may be a better practical fit than a general-purpose multimodal model
• –The post points to a common local-AI tradeoff: aggressive quantization and narrow prompts often matter more than chasing the strongest model family