✓Direct Answer

Yes. Oyma connects natively to Ollama running on your Mac (`http://localhost:11434`). You can use open models like Llama 3, Mistral, and Gemma to summarize meetings and query notes with zero data leaving your machine.

Private AI Synthesis Without Remote APIs

While on-device Whisper converts spoken audio into written text, knowledge workers also need intelligent synthesis: executive summaries, action item checklists, and technical key takeaways. Relying on cloud APIs (like OpenAI or Anthropic) introduces recurring costs and sends confidential notes to external servers.

Oyma pairs seamlessly with Ollama, the premier tool for running open large language models locally on macOS.

Setting Up Ollama with Oyma

Connecting Oyma to Ollama takes under two minutes:

  1. Install Ollama: Download and install Ollama from ollama.com on your Mac.
  2. Pull Your Preferred Model: Open Terminal and download a state-of-the-art open model: 
    ollama pull llama3:8b
# Or for larger unified memory systems:
ollama pull mistral-nemo:12b
  3. Configure Oyma: In Oyma Preferences > AI Providers, select Ollama (Local).
  4. Automatic Model Discovery: Oyma queries http://localhost:11434/api/tags and populates your installed models in the dropdown menu.
+-----------------------------------------------------------+
|                   Local Mac Workstation                   |
|                                                           |
|   [Oyma Meeting Note]  <==== HTTP Localhost ====>    |
|                                                           |
|                  [Ollama Inference Engine]                |
|                             |                             |
|              [Llama 3 8B / Apple Silicon Metal]           |
|                                                           |
|                  (ZERO EXTERNAL NETWORK TRAFFIC)          |
+-----------------------------------------------------------+

Optimal Open Models for Meeting Summaries

  • Llama 3 8B / 3.1 8B: Exceptional structured reasoning, fast token generation (~45 tokens/sec on M2 Pro).
  • Mistral 7B / Nemo 12B: Excellent nuance and concise executive briefing capabilities.
  • Qwen 2.5 7B / 14B: Superb multilingual performance and precise code extraction.

Unified Memory Advantages for Local LLMs

Apple Silicon’s unified memory architecture is uniquely suited for running large language models:

  • Shared High-Bandwidth RAM: The CPU, GPU, and Neural Engine access the same pool of high-speed memory, eliminating PCIe bus bottlenecks.
  • Running 7B to 14B Parameter Models: Macs with 16 GB or 36 GB of unified memory can run models like Llama 3 8B or Mistral 7B at 40+ tokens per second.
  • Zero API Ingestion Risk: Local models execute entirely offline without transmitting your meeting transcripts or notes across external network interfaces.

Local Model Freedom and Offline Autonomy

With Ollama running locally, you can switch between models like Llama 3, Mistral, and Gemma with a single click. You maintain complete control over model sizes, quantization levels, and system prompt parameters without paying API fees.

Because Ollama runs locally on your Mac’s unified memory, inference speeds easily exceed 40 tokens per second on M2 Pro and newer chips, producing full meeting summaries in under 15 seconds.

Local Inference Technical Invariants

  • Host Connection: Communicates via localhost HTTP (http://localhost:11434) with zero external traffic.
  • Supported Model Families: Llama 3, Mistral, Gemma, Phi-3, Qwen 2.5, and Command-R.
  • Hardware Acceleration: Metal acceleration leverages Apple Silicon unified memory at up to 800 GB/s bandwidth.
  • Token Throughput: Exceeds 40 tokens per second on M-series chips, summarizing meetings in seconds.

Related Questions & Architecture Guides