Local intelligence
Running AI locally
Four model weights live on the Mac, exposed through six Ollama tags. Three of them form a pressure-aware specialist router; the fourth stays available for experiments.
- Model weights
- 4
- Router lanes
- 3
- Custom MCP tools
- 6
- Installed weights
- 18.2 GB
The model roster
Small models, routed by the shape of the job.
Llama 3.2 3B
Extraction, classification, tagging, and short rewrites
Qwen2.5-Coder 7B
Code and structured transformations
Qwen3.5 9B
Research synthesis, long documents, and ambiguous work
L3 8B Stheno v3.2
A separate quantized model kept outside the production router
The token router
Fast work stays fast; harder work earns more context.
- Fast laneLlama 3.2 3B · 4K context
- Coder laneQwen2.5-Coder 7B · 8K context
- Strong laneQwen3.5 9B · 16K, or 32K while exclusive
- Admission controlUp to three simultaneous jobs with memory-pressure gates
The current 30-day routing log records 55 lane selections with no routing failures. New inference pauses under elevated memory pressure or when available memory falls below 2 GB.
MCP and local tools
Two custom MCP servers: local_qwen for bounded local work and node_repl for persistent JavaScript.
- Local task delegation
- Public-web research
- Headline extraction
- Document parsing
- Local file search
- Usage reporting
The stack speaks MCP over stdio, native Ollama HTTP, and an OpenAI-compatible chat-completions API. SQLite backs routing state, caching, and content-free usage metrics.
Internet access
Local search first, browser only when needed.
A self-hosted SearXNG container handles public search. Static page retrieval is preferred; one bounded headless-browser slot can step in when ordinary fetching fails. Results are cached, citations are retained, and research synthesis returns to the strong local model. There is no paid search API in this path.
Also in the lab
Photo Memory
A native Swift 6 macOS prototype for read-only Photos inventory, local SQLite metadata, and model-versioned vector search. The app is deliberately offline: Photos stays untouched, previews are not persisted, and production semantic search remains an experiment rather than a finished claim.