Meeting summaries with a local LLM: Ollama and LM Studio on the Mac – also on macOS 14 and 15
Key takeaways
- Voxdrop 2.3 can write meeting summaries and to-dos with a model you run yourself in Ollama or LM Studio – not only with Apple Intelligence.
- That brings summaries to macOS 14 and 15 and to anyone who keeps Apple Intelligence off. Voxdrop only accepts addresses on the same Mac, so the text stays local.
- The trade-off, measured: Gemma 3 4B needed 14–27 seconds per text block on an M3 with 24 GB, about four times as long as Apple's model. By default it summarizes after the call.
Voxdrop transcribes calls on your Mac. For the summary, the short version and the to-dos it needs a language model, and up to version 2.2 that meant Apple's on-device model, which requires macOS 26 or later and Apple Intelligence turned on. Version 2.3 adds a second option: a local server you run yourself, with Ollama or LM Studio. This post covers the setup, what we measured and where it falls short.
When do you need a local model for meeting summaries?
Apple's Foundation Models framework – the on-device model that apps can use – starts with macOS 26 and only runs when Apple Intelligence is turned on (Macs with M1 or later). A local server makes sense if:
- your Mac runs macOS 14 Sonoma or 15 Sequoia (meeting recording needs 14.4 or later),
- you keep Apple Intelligence off,
- or you want to pick the model yourself.
| Apple (on this Mac) | Ollama | LM Studio | |
|---|---|---|---|
| Requires | macOS 26+, Apple Intelligence on | macOS 14+ | macOS 14+, Apple Silicon |
| Setup | None | App + one command | App + start the server |
| Model | Apple's model | Your choice | Your choice |
| Where the text goes | Stays on the Mac | Stays on the Mac | Stays on the Mac |
| Cost | None | None | None |
How do I connect Ollama to Voxdrop?
- Install Ollama from ollama.com (the Mac app needs macOS 14 or later).
- Download a model in Terminal, for example
ollama pull gemma3:4b(about 3.4 GB). - In Voxdrop: Settings → Meeting Detection → AI for summaries and to-dos → Local server → Ollama. The address is filled in:
http://127.0.0.1:11434. - Click Load models and pick yours. The next meeting is summarized with it; for older meetings, use Create summary in the meeting.
And with LM Studio?
LM Studio (macOS 14 or later, Apple Silicon) has a built-in server: open the Developer tab and switch on Start server, or run lms server start in Terminal. It listens on port 1234 by default. In Voxdrop, choose Local server → LM Studio, then Load models. If LM Studio's server isn't running, Voxdrop tells you so instead of failing silently.
How fast is a local model in practice?
We ran Voxdrop's block-by-block processing with Gemma 3 4B in Ollama on an M3 MacBook with 24 GB, against copies of six real meetings of 2 to 34 minutes (we only looked at the numbers, not the content):
| Measured (M3, 24 GB) | Apple's model | Gemma 3 4B in Ollama |
|---|---|---|
| Seconds per text block | 4–6 | 14–27 |
| Failed blocks, 6 real meetings | 0 | 0 |
| Memory used by the model | managed by macOS | 2.9 GB |
| Download | Part of macOS | ~3.4 GB |
A 30-minute call therefore takes a few minutes until the summary is ready, instead of seconds. The to-dos are quicker than that, because Voxdrop only sends the passages around commitments and requests to the model, not the whole transcript. How that works is in the post on meeting to-dos.
Why does Voxdrop only accept addresses on this Mac?
Voxdrop's promise is that your meetings don't leave your Mac. A “local server” field that accepted any URL would be one typo away from sending transcripts somewhere else. So Voxdrop only accepts loopback addresses – 127.0.0.1, localhost or ::1 – and refuses everything else. Ollama and LM Studio both listen there by default.
A detail for developers: Ollama's context window
If you build something similar: Ollama's default context window is small – 4,096 tokens according to its FAQ; newer versions scale it with available memory. The docs don't say what happens to longer input, and reports in Ollama's issue tracker describe it being cut without an error. Voxdrop therefore uses Ollama's native /api/chat endpoint, sets num_ctx to 8,192 per request, keeps text blocks around 2,500 characters and compares the returned prompt_eval_count with the expected size. If the input was cut, the block counts as failed – better than a summary of half a transcript. LM Studio is reached through its OpenAI-compatible /v1/chat/completions.
Honest caveats
- Slower. Several minutes for a long call with a 4B model, and live summaries are off by default.
- Memory and disk. Gemma 3 4B is about 3.4 GB on disk and used 2.9 GB of memory in our test; bigger models need more.
- Quality depends on the model. Small models sometimes ignore the answer format; Voxdrop then skips that block rather than guessing.
- Dictation modes stay with Apple. Rewriting dictated text still uses Apple's model.
- Model licenses are the model's. Gemma 3, for example, comes under Google's Gemma Terms of Use.
(Disclosure: Voxdrop is our app.) Voxdrop isn't alone here: MacWhisper, Meetily, Minutes and anarlog can also use Ollama for summaries. Why on-device matters for confidential conversations is covered in local voice-to-text for Mac.
FAQ
Can I summarize meetings with Ollama or LM Studio instead of Apple Intelligence?
Yes, since Voxdrop 2.3. Under Settings → Meeting Detection → "AI for summaries and to-dos", choose Local server, then Ollama or LM Studio and a model. Summaries, short versions and to-dos then come from that model.
Do meeting summaries work on macOS 14 Sonoma or 15 Sequoia?
With a local model, yes. Apple's on-device model for apps needs macOS 26 or later with Apple Intelligence turned on. With Ollama or LM Studio, Voxdrop summarizes on macOS 14.4 and later, the minimum for meeting recording.
Does my meeting text leave the Mac with a local model?
No. Voxdrop only accepts server addresses on the same Mac – 127.0.0.1, localhost or ::1 – and refuses anything else, so a transcript cannot be sent to a remote server by mistake.
Which local model should I use for meeting summaries?
We tested Gemma 3 4B in Ollama: no failed blocks on six real meetings and 2.9 GB of memory, but about four times slower than Apple's model. Larger models can be better and need more memory. Judge the quality on your own meetings.
Does the local model summarize while the call is still running?
Not by default. Running a language model next to live transcription competes for the same chip and memory, and we have not measured that yet. The summary starts when you hang up; you can turn on "Summarize during the call" yourself.
Read next
- Meeting action items on the Mac, without a bot in the call
- Ask Claude about your meetings: a local MCP server
- Local, offline voice-to-text for Mac
Get Voxdrop
Voxdrop is on-device voice-to-text for any Mac app — hit a hotkey, speak, and the text appears at your cursor. Local Whisper, €9.99 Early-Bird (first 500 buyers) or €14.99 Standard, one-time, no subscription. License key right after checkout, 14-day money-back.