Memory consumption and optimization

Apple silicon Macs utilize unified memory shared between the CPU, GPU, and Neural Engine. Running machine learning models locally delivers blistering speed, but requires thoughtful memory management on 8 GB and 16 GB machines.

Memory footprints of local AI models

Component	Active RAM Footprint	Recommended Mac Config
Oyma Base Application	~180–250 MB	Any Mac
Whisper Base Model	~320 MB	8 GB+
Whisper Small Model	~850 MB	8 GB+
Whisper Large v3 Turbo	~2.1 GB	16 GB+
Ollama Llama 3.2 3B	~3.5 GB	8 GB+
Ollama Llama 3.1 8B	~8.0 GB	16 GB+

Optimization strategies for 8 GB Macs

1. Model Pairing: Pair Whisper Small with Llama 3.2 3B. Total memory consumption remains under 4.5 GB, leaving 3.5 GB free for macOS and productivity apps.
2. Sequential Execution: By default, Oyma runs transcription first, unloads audio buffers, and then invokes Ollama for summary generation.
3. Auto-Unload Idle Models: In Settings › AI › Memory, set Model Idle Timeout to 5 minutes to return RAM to the system immediately after synthesis.

Monitoring memory in macOS Activity Monitor

Open Activity Monitor (⌘Space › Activity Monitor) and select the Memory tab:

• Green Memory Pressure: Optimal. The system is executing at full speed without swapping.
• Yellow Memory Pressure: Acceptable during intensive batch processing.
• Red Memory Pressure: The system is memory-constrained. Reduce model sizes to restore peak performance.

Related guides

• Model benchmarks in Whisper benchmarks on Apple silicon.
• Ollama setup in Ollama local setup and model configuration.
• Batch processing in Batch audio and video transcription.