In April, I called running AI on a Mac Mini the biggest cope in tech. LFM3 is making me rethink that. It's a 1 GB model from Liquid AI, trained specifically for agent tools like OpenClaw and Hermes. In Liquid's own agent test, it beats a Qwen model five times its size, and it runs offline on your phone. Don't use it to write code (for that, I still pay for the big models). But a lot of what my agents do all day is small, boring stuff like tagging support tickets and summarizing logs. Run it in Ollama or LM Studio and hand it the boring jobs. Every job it handles comes off your API bill.