Five local models, 6.8 GB of weights: my open-source Mac meeting notetaker

preview.redd.it/xs01ds930nsh1.png Back in July I shared LokalBot here. It's a free, open-source Mac app that records your meetings and keeps a daily summary of your activity, all on-device. 0.9.2 came out today. Since July I've benchmarked every model in it and swapped most of the defaults for smaller ones. The whole stack is now 6.8 GB: Qwen3-ASR 1.7B (MLX, 8-bit): transcription Nemotron 3 (Core ML): who spoke when Qwen3.5 4B Q4_K_M (llama.cpp): notes and action items Harrier 0.6B Q8_0: search embeddings LFM2.5 1.2B Q4_K_M: autocomplete in any app Apple Vision : screen OCR (opt-in) A few numbers from my M4 Max (48 GB): 26-min meeting to finished notes in 33 s warm , ~85 tok/s decode Speaker error went from 43.4% to 14.6% DER on AMI. That's against my old pyannote setup, so it says more about my config than about pyannote. Autocomplete p95 went from 1.83 s (Gemma 4 E4B) to 0.49 s There's also a read-only MCP server and CLI , off by default, so Claude Code or any other MCP client can pull context from your meetings. I don't have any 16 GB or other M series numbers yet. If you've got one of those, especially M5 or M6 I'd love to see what you get. I also tried MiniCPM5 2B for notes. It was smaller and faster, but it got stuck repeating itself on one summary and assigned action items to the wrong person. I kept Qwen3.5 4B as the default as saving a few seconds wasn’t worth getting who agreed to do what wrong. Repo What's new in 0.9.2 All benchmark numbers with sources