Local AI setup
Hyprcore can run a language model on your Mac for dictation cleanup, meeting summaries, and knowledge chat—fully on-device, no cloud round-trip. It uses Ollama, a free third-party app that runs models locally.
You set this up from a dialog on the Chat (Knowledge) page—not during onboarding. When no AI model is configured yet, the Chat page shows a Set up local AI card. Click it to open the local AI setup dialog. (Prefer the cloud? The same card has a Use cloud instead option that sends you to Settings → Models.)
Note: Ollama runs as a background service on
http://localhost:11434. Quitting it from the menu bar stops Hyprcore's local LLM features until you start it again.
What the dialog does
The dialog walks two steps and polls Ollama every few seconds so it updates as you go:
Install and run Ollama. It checks whether Ollama is installed and running.
Download a model. Once Ollama is running, you pick a model and download it.
When a model is ready, Hyprcore wires Ollama as the provider for dictation post-processing, meeting summaries, and knowledge chat.
Install Ollama
Download the macOS app. Click Install Ollama in the dialog, or go to ollama.com/download/mac.
Open the .dmg and drag Ollama into Applications. Standard macOS install. Approve the Gatekeeper prompt the first time.
Launch Ollama. Open it from Applications or Spotlight. The first launch starts the background service and adds a llama icon to the menu bar. Leave it running.
Return to Hyprcore. Within a few seconds the dialog flips Step 1 to a green check. You don't need to restart Hyprcore.
Choose a model
Once Ollama is running, the dialog offers three models. Hyprcore highlights a recommended one based on how much RAM your Mac has, and flags a model with an amber warning if it's likely too large for your machine.
Model | Size on disk | Recommended RAM | Knowledge chat | Best for |
|---|---|---|---|---|
Qwen 3.5 4B (default) | ~3.4 GB | ~5 GB free | Full mode (tool-capable) | The best all-round default. Cleanup, summaries, and knowledge chat. |
Qwen 3.5 9B | ~6.6 GB | ~9 GB free | Full mode (tool-capable) | Stronger answers and multi-step research on 16 GB+ Macs. |
Gemma 3 1B | ~815 MB | ~2 GB free | Light mode only | A tiny offline fallback for almost any Mac. |
Tool-capable models unlock Full mode knowledge chat, where the Agent can run tools and do multi-step research across your vault. Gemma 3 1B runs knowledge chat in Light mode only.
Pick one and click Download. A progress bar shows a live percentage. When the pull finishes, Hyprcore sets the model as your provider and you're done.
If you'd rather pull manually from a terminal:
ollama pull qwen3.5:4b
# or
ollama pull qwen3.5:9b
# or
ollama pull gemma3:1b
Hyprcore picks up the model on its next status check.
Tip: The same dialog can also download the optional speaker-identification (diarization) models used to label who said what in meeting transcripts.
What Hyprcore configures
When setup completes, Hyprcore writes these settings for you:
Settings → Models → Post-processing → Provider: Ollama
Settings → Models → Post-processing → Model: the model you chose
Settings → Meetings → Processing provider: Ollama
You can change any of these later in Settings → Models.
Troubleshooting
Step 1 stays gray after I install Ollama
Ollama has to be running, not just installed. Open it from Applications. You should see a llama icon in the menu bar. The dialog polls every 3 seconds, so it should turn green within a few seconds of launch.
If the menu bar icon is there but Step 1 is still gray, your firewall or a VPN may be blocking localhost. Try curl http://localhost:11434/api/tags from a terminal—a working Ollama responds with a JSON list (possibly empty).
The model download fails partway through
Ollama writes partial pulls into ~/.ollama/models. To retry cleanly:
ollama rm qwen3.5:4b
ollama pull qwen3.5:4b
Then click Download in Hyprcore again. Common causes are flaky Wi‑Fi and corporate proxies that interrupt long downloads—a wired connection helps.
Port 11434 is already in use
Another tool (a previous Ollama install, a dev server, or LM Studio's compatibility layer) is holding the port. Find it with:
lsof -i :11434
Quit the conflicting process, then relaunch Ollama. Hyprcore expects Ollama at the default http://localhost:11434.
My Mac runs hot or feels slow during a summary
Qwen 3.5 9B uses significantly more RAM and CPU than the smaller models. On 8 GB Macs, switch to Qwen 3.5 4B, or Gemma 3 1B for the lightest footprint, in Settings → Models → Post-processing. You can also route meeting summaries through Hyprcore Cloud or a BYOK provider for higher quality.
I already have other Ollama models pulled
Anything you've pulled in Ollama shows up automatically in Settings → Models → Post-processing → Model when the provider is Ollama. See Custom and local models for the full list.
I don't want to install Ollama at all
Choose Use cloud instead on the setup card, or Skip in the dialog. Hyprcore uses Hyprcore Cloud (powered by your plan tokens) by default, and you can switch to BYOK Anthropic, OpenAI, Google, Groq, or LM Studio in Settings → Models at any time. See Post-processing LLMs for the full provider list.
Still stuck
Live chat. The Helpin chat widget in the web dashboard.
Discord. discord.gg/UEvrjftB.
Email. support@hyprcore.ai with your Ollama version (
ollama --version) and Hyprcore logs attached.
Was this article helpful?

