Meeting summaries with a local LLM: Ollama and LM Studio

Autorec already records and transcribes on your machine. The summary is the one step that can leave it, because a summary needs a language model and most people use a hosted one.

It doesn’t have to be. Autorec talks to any OpenAI-compatible endpoint, and Ollama and LM Studio both expose one on localhost. Point Autorec there and nothing leaves the machine at any stage. This post walks through both, then covers which model to pick for 8, 16 or 32 GB of RAM and why long meetings need a bigger context window than the default.

What actually gets sent

Autorec sends two things to the endpoint: your summary prompt as the system message, and the whole transcript as the user message. It sends no audio or video. (Autorec cuts transcripts beyond roughly 400,000 characters, about eight hours of speech, with a ”[Transcript truncated]” marker.) With a local server that request goes to http://localhost:11434/v1 (Ollama) or http://localhost:1234/v1 (LM Studio) and stops there.

Option 1: Ollama

1. Install Ollama from ollama.com/download. On Linux, the install script is curl -fsSL https://ollama.com/install.sh | sh. On macOS and Windows it’s a normal installer.
2. Pull a model in a terminal, for example ollama pull qwen3.5:9b. Model picks are below.
3. Check it’s running: Ollama listens on 127.0.0.1:11434 by default, and ollama ps shows what’s loaded.
4. In Autorec, open Settings and scroll to AI Summary. Under Provider, choose Ollama (local) from the Local servers group. The API Base URL fills in as http://localhost:11434/v1.
5. Leave API Key empty. The field reads “Not needed for local servers”.
6. Set Model to the exact tag you pulled, such as qwen3.5:9b. The preset suggests llama3.1, which only works if you pulled that one.
7. Click Save Settings.

Option 2: LM Studio

1. Install LM Studio from lmstudio.ai and download a model from its built-in search.
2. Start the server. In LM Studio’s Developer tab, switch on Start server, or run lms server start in a terminal. It serves on port 1234.
3. In Autorec, Settings, AI Summary, choose LM Studio (local) as the Provider. The base URL becomes http://localhost:1234/v1.
4. Leave API Key empty.
5. The LM Studio preset has no default model, so type the model identifier exactly as LM Studio shows it.
6. Click Save Settings.

LM Studio is the easier choice if you’d rather click than type commands. Ollama is lighter and sits in the background without a window.

Test it on one recording

Open a recording with a transcript in your library and click Generate Summary. If it works, the title and summary appear once the model has finished, which depends on your hardware. Once you’re happy, switch on Auto-summarize after transcription so each new recording gets one as soon as it’s transcribed.

Picking a model

Two numbers matter: the model’s file size, which has to fit in RAM (or VRAM) with room left for your meeting app and Autorec, and its context window, which decides how much transcript it can read at once.

As of October 2026, these are tags on ollama.com/library. Sizes are the download sizes listed there for the default quantisation.

Your RAM	Model	Download size
8 GB	qwen3.5:4b	about 3.3 to 4 GB
16 GB	qwen3.5:9b or gemma4:e4b	about 6.6 to 7.6 GB
32 GB	gemma4:26b or qwen3.5:27b	about 16 to 20 GB

These are starting points from a model list, not benchmark results. To find out which one writes the best meeting summary for you, run your own transcripts through two of them. The model libraries move fast, and the names above will age. Check the library page before pulling.

• 8 GB is tight. You may be running a call, a screen capture, a Whisper pass and a model on the same machine. A 4B model works, but expect rougher summaries and keep other apps closed while it runs.
• Transcription and summarising don’t have to overlap. Autorec transcribes first and summarises after, so the model only needs memory once the recording is over.
• A GPU helps a lot. On CPU alone a 9B model is usable but slow. A one-hour transcript can take several minutes.
• Smaller models drift on long input. If a summary misses the second half of a meeting, try a larger model before blaming the prompt.

Long meetings: raise the context length

This is the setting that catches people out. A one-hour meeting is roughly 9,000 words of transcript, which comes to somewhere around 12,000 tokens or more. A model can only read as much as its context window holds.

Ollama’s docs say the default context length depends on your VRAM: 4k tokens under 24 GiB, 32k at 24 to 48 GiB, 256k at 48 GiB or more. So on a typical laptop or a mid-range GPU the default is 4,096 tokens, and a long transcript won’t fit: the transcript, the prompt and the reply all have to fit in that window. You may not get a clear error, just a summary that covers part of the meeting.

To raise it, set OLLAMA_CONTEXT_LENGTH before the server starts:

OLLAMA_CONTEXT_LENGTH=32768 ollama serve

If Ollama runs as a service, put the variable in the service environment instead. On Linux that’s systemctl edit ollama.service with an Environment="OLLAMA_CONTEXT_LENGTH=32768" line under [Service], then a restart. The Ollama desktop app also has a context length slider in its settings. Afterwards, run ollama ps while a summary is generating and check the CONTEXT column to confirm the new value took effect.

The Extra request fields (JSON) box in Autorec’s AI Summary settings won’t help here. Ollama’s OpenAI-compatible endpoint has no field for context size, and its docs say so. The alternative to the environment variable is a Modelfile with PARAMETER num_ctx 32768, built with ollama create into a new model name that you enter in the Model field.

A bigger context costs memory. 32k tokens is enough for most hour-long meetings and is a sensible first try. If ollama ps shows the model split between GPU and CPU, you’ve asked for more than your VRAM holds, so lower it or use a smaller model.

In LM Studio, context length is a load setting for each model. You can set it when you load the model and save it as that model’s default from the gear icon in My Models. Pick a value that covers your longest meetings, and keep an eye on memory use.

Also check Max summary length (tokens) in Autorec’s AI Summary settings. It caps the length of the reply, not the input. The help text says 300 is a short paragraph and 1000 is roughly a page.

What you end up with

The recording is an .mp4 on your disk. The transcript is made by Whisper on your CPU or GPU. The summary comes from a model running on localhost. You create no account, and no vendor holds a copy of anything. If a client asks where their call went, you can answer in one sentence. For a wider look at the same idea, see transcribing meetings locally with Whisper.

The cost is quality and speed. A 9B model on a laptop won’t match a frontier model on a hosted API. If that gap matters, there are middle paths. Summaries through an EU-hosted endpoint keep data under GDPR, and OpenRouter gives you hosted models at pay-per-use prices.

Once it works, the prompt is next. Editing the summary prompt is how you get decisions, action items and owners instead of a generic recap.

Troubleshooting

• Connection refused. The server isn’t running. For Ollama, check ollama ps or restart the service. For LM Studio, make sure Start server is on.
• Model not found. The Model field has to match the tag or identifier exactly, including the size suffix.
• A summary that ignores the end of the meeting. The context window is too small. Raise it as described above.