Agent Synth · Catalog
Hand-picked local models, sized for Apple Silicon. We update this list as the landscape shifts. The menu-bar agent uses this catalog to recommend swaps that fit your machine.
qwen2.5-coder:1.5b
Tiny coding model for fast autocomplete on 8 GB Macs.
Best as a draft/autocomplete model — not for long-form code generation.
Community-curated · added 2026-05-28
qwen2.5-coder:7b
Solid coding default for 16 GB+ Macs; strong instruction following.
Community-curated · added 2026-05-28
qwen2.5-coder:14b
Stronger coding — fits 24 GB+ Apple Silicon at default quant.
Supersedes
Community-curated · added 2026-05-28
qwen2.5-coder:32b
Frontier-tier local coding model — fits M-series Max/Ultra.
Supersedes
Community-curated · added 2026-05-28
llama3.2:3b
Small, fast chat — viable on 8 GB Macs and as a low-latency draft model.
Community-curated · added 2026-05-28
qwen2.5:7b
Strong general chat at 7B. Use Coder variant for code-heavy work.
Community-curated · added 2026-05-28
llama3.1:8b
Older default — still solid for chat/RAG with the broadest runtime support.
Community-curated · added 2026-05-28
qwen2.5:14b
Capable general chat — preferred over 7B once memory allows.
Supersedes
Community-curated · added 2026-05-28
deepseek-r1-distill-qwen:7b
Compact reasoning model — visible chain-of-thought useful for code planning + math.
Verbose by default — outputs <think> blocks. Strip or surface them depending on the surface.
Community-curated · added 2026-05-28
deepseek-r1-distill-qwen:14b
Mid-tier reasoning — stronger on multi-step problems than the 7B at the memory cost.
Supersedes
Community-curated · added 2026-05-28
nomic-embed-text
Default embedding model for local RAG. Tiny footprint, broad runtime support.
Community-curated · added 2026-05-28
mxbai-embed-large
Alternative embedding — competitive quality, slightly larger vectors than Nomic.
Community-curated · added 2026-05-28
llama3.2-vision:11b
Image-aware chat — for screenshot Q&A, OCR-like extraction, and visual reasoning.
Ollama-only at default quant today. MLX support exists for separate weights.
Community-curated · added 2026-05-28
whisper-large-v3
Gold-standard local transcription — strong multilingual + low WER.
Typically loaded via whisper.cpp or mlx-whisper, not via Ollama. Used by Hermes for STT.
Community-curated · added 2026-05-28
distil-whisper-large-v3
~6× faster than Whisper Large v3 with a small accuracy trade-off — good for live transcription.
Supersedes
Optimized for English. Pair with Whisper Large for multilingual.
Community-curated · added 2026-05-28