Update, October 2026: several default models have changed since this post; the current ones are on the benchmarks page.
What it does
- Meetings. Detects Zoom, Teams, Meet, and other calls and records you and the other side on two synchronized tracks, with no bot joining the call. Transcription and the summary run on your Mac as soon as the meeting ends.
- Dictation. Hold ⌥ Space, talk, and release. It's transcribed locally and pasted at the cursor, and the audio is deleted afterwards.
- Autocomplete. Ghost-text suggestions in almost any Mac text field; Tab accepts. Opt-in.
- Recall. Ask “what did we decide about this?” and get the answer from your own library, with the source meeting cited and the exact moment ready to replay. Full-text and semantic search cover everything.
- Day timeline. An optional private timeline of where your time went, with a daily digest.
New in 0.5
- Dreaming. LokalBot can compile a finished day overnight into a morning brief and structured work memory.
- Faster autocomplete. New installs get the 0.73 GB LFM2.5 1.2B model, and accepting suggestions is more reliable across native apps, Chromium, and input methods.
- Answers in Quick Recall. Quick Recall can answer right in its compact window instead of always opening the full Ask view.
- Scheduled day digests, with a prompt you can customize.
The models in daily use
One setup, on a 48 GB M4 Max, all through the bundled llama.cpp with full Metal offload:
| Job | Model | Quantization | Size | Speed |
|---|---|---|---|---|
| Transcription | Granite Speech 4.1 2B | Q4_K_M + F16 projector | 2.3 GB | Faster than real time |
| Summaries and chat | Qwen3.5 4B | Q4_K_M | 2.7 GB | About 100 tokens/s |
| Autocomplete | Gemma 4 E4B | UD-Q5_K_XL | 6.7 GB | About 78 tokens/s |
| Embeddings | Qwen3-Embedding 0.6B | Q8_0 | 0.6 GB | — |
| Diarization | pyannote community-1 | Core ML | 0.1 GB | — |
The full list of changes is in the 0.5.0 release notes.