> ## Documentation Index
> Fetch the complete documentation index at: https://docs.stenoai.co/llms.txt
> Use this file to discover all available pages before exploring further.
# Summarization models
> Steno generates meeting summaries using local Ollama models by default, with optional cloud model support. Choose a model based on your hardware, privacy requirements, and the quality of notes you need.
After transcription, Steno passes the transcript to a language model to generate structured notes: a summary paragraph, key topics, and action items. By default, this uses a local model running via [Ollama](https://ollama.com) entirely on your Mac. You can optionally configure a cloud model (OpenAI, Anthropic, AWS Bedrock, or a custom endpoint) — see the section below.
## Available models
| Model | Size | Notes |
| - | - | - |
| `gemma4:e2b-it-qat` | \~4.3GB (\~6.5GB on Apple Silicon) | Default. Fast, 128K context. |
| `gemma4:e4b-it-qat` | \~6.1GB (\~8.8GB on Apple Silicon) | Higher quality, modest footprint. |
| `qwen3.5:9b` | \~6.6GB | Strong at structured output and action items. |
| `gemma4:12b-it-qat` | \~7.2GB (\~7.7GB on Apple Silicon) | 256K context, best for long meetings. |
| `gpt-oss:20b` | \~14GB | Highest quality, reasoning. |
On Apple Silicon, the three Gemma 4 models are automatically pulled as MLX/NVFP4 builds instead of the base GGUF -- a meaningful speed win, at a larger download size (shown above).
`gemma4:e2b-it-qat` (Gemma 4 E2B) is the default and handles most meetings well. `llama3.2:3b` is still available for existing users pinned to it, but is deprecated in favor of the Gemma 4 lineup and hidden from the default model picker. Larger models produce better-structured, more detailed notes; the trade-off is processing time and disk space.
## How to change models
1. Open **Settings → AI**
2. Select a model from the list
3. If the model is not yet downloaded, Steno will download it via Ollama
The new model is used for all future recordings. You can re-summarize an existing recording from its detail view using any model.
## Choosing the right model
**For daily use:** `gemma4:e2b-it-qat` (default) works well for most meetings. Notes are clear and concise without requiring significant processing time.
**For important meetings:** `qwen3.5:9b` or `gemma4:12b-it-qat` produce more detailed notes with better-identified action items and more accurate key topics. `gemma4:12b-it-qat` offers a 256K context window, which helps with long meetings.
**For the highest quality:** `gpt-oss:20b` produces the best local output, at the cost of more disk space and slower processing.
## Using a cloud model
Steno optionally supports OpenAI, Anthropic, AWS Bedrock, or a custom API endpoint as an alternative to a local model. If configured, your transcript and any typed notes (not your audio) are sent to that API for summarization.
To configure a cloud model:
1. Open **Settings → AI**
2. Select **Cloud API**
3. Choose the provider and enter your credentials
Cloud models are off by default. If you work with confidential recordings, use a local model.
***
`qwen3.5:9b` is particularly good at identifying and formatting action items from meeting transcripts. If structured output is important, it is worth the extra size over the default `gemma4:e2b-it-qat`.
The default setup (Parakeet + `gemma4:e2b-it-qat`) requires approximately 7GB. The largest model (`gpt-oss:20b`) requires \~14GB. Ollama models are stored in `~/.ollama/models/`.
Yes. Run `ollama rm [model-name]` in Terminal to remove a model and free the disk space. You can re-download it later from within Steno.
Stick with `gemma4:e2b-it-qat` (the default) -- it has the smallest footprint of the curated lineup. The larger models need more RAM to hold the model weights, so on an 8GB Mac they are more likely to cause swapping, which slows both summarization and the rest of your system.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.