ollama

name	ollama
description	Use when running Ollama commands on <MAC_HOST> () — listing/pulling/showing/running models, inspecting templates, switching context. Captures the right SSH invocation, binary location, env vars, and the gotchas that have bitten before. Pair with /docker-switch for Docker Model Runner side-by-side.
category	execution
runtimes	["claude"]
pii_safe	true

Ollama on Mac

Mac runs Ollama as a desktop app on <PRIVATE_IP>:11434. From any other machine, ssh <MAC_HOST> <ollama-command>. From Mac, just ollama. <USER_NAME>has corrected this multiple times — don't keep guessing paths.

Binary location & PATH

Path	What	Notes
/Applications/Ollama.app/Contents/Resources/ollama	Real binary	Bundled with desktop app
/usr/local/bin/ollama	Symlink → real binary	Works in interactive shells
/opt/homebrew/bin/ollama	Symlink → /usr/local/bin/ollama	Required for non-interactive SSH because Mac's default SSH $PATH does NOT include /usr/local/bin

If ssh <MAC_HOST> ollama returns "command not found", the /opt/homebrew/bin symlink got nuked. Recreate:

ssh <MAC_HOST> 'ln -sf /usr/local/bin/ollama /opt/homebrew/bin/ollama'

Daily commands (from anywhere)

ssh <MAC_HOST> ollama list                       # what's pulled
ssh <MAC_HOST> ollama ps                         # what's loaded in Metal RAM right now
ssh <MAC_HOST> ollama pull <model>               # pull a model
ssh <MAC_HOST> ollama show <model>               # capabilities, parameters, license
ssh <MAC_HOST> ollama show --template <model>    # raw chat template (often `{{ .Prompt }}` passthrough)
ssh <MAC_HOST> ollama show --modelfile <model>   # full Modelfile (params, system, FROM)
ssh <MAC_HOST> ollama rm <model>                 # delete a model
ssh <MAC_HOST> ollama run <model>                # interactive REPL — usually skip, prefer pi/Cline/opencode

API endpoint

OpenAI-compatible:

• LAN: http://<PRIVATE_IP>:11434/v1
• Local on Mac: http://127.0.0.1:11434/v1
• API key: literally ollama (no real auth)

Generic also has OLLAMA_HOST=<PRIVATE_IP>:11434 env so its local Ollama CLI talks to Mac.

Loading status / TTL

• Default keepalive: ~4 minutes idle → model auto-unloads from Metal
• First request after unload reloads (slow, model-dependent — 10–30s for 30GB)
• One model at a time on Metal compute; loading a second will fail or evict the first
• Practical Mac Metal compute graph limit: ~25GB safe, ~45GB risky, 60GB+ fails (despite 128GB unified RAM)

Auto-sync pi config with pulled models

pi-sync-ollama (deployed at ~/.local/bin/pi-sync-ollama on <WORKSTATION_HOST>/generic, ~/.local/bin/pi-sync-ollama on Mac) hits Ollama's /v1/models, builds a fresh ~/.pi/agent/models.json for every pulled model, and applies family-specific capability flags (Qwen 3.5/3.6 → compat.thinkingFormat:"qwen" + reasoning:false to avoid loops; Qwen3-VL → image input; Devstral → JSON tools; etc.). Backs up the old config to models.json.bak.<timestamp>.

# Standard workflow after pulling a new model:
ssh <MAC_HOST> ollama pull <new-model>
pi-sync-ollama                  # on whichever box you're working on
ssh <INFERENCE_HOST> ~/.local/bin/pi-sync-ollama   # if you want generic updated too
ssh <MAC_HOST> ~/.local/bin/pi-sync-ollama       # likewise for Mac

# Dry-run / preview what config would be written:
pi-sync-ollama --print

# Override endpoint (e.g. for a remote Ollama):
OLLAMA_BASE_URL=http://other-host:11434/v1 pi-sync-ollama

The sync script's family heuristics live in its case statement — extend it when new model families show up. Default behavior for unrecognized models is conservative (text-only, 32K context, reasoning off).

Pulling from Hugging Face GGUFs

Ollama can pull HF GGUF directly using the hf.co/<repo> syntax:

ssh <MAC_HOST> ollama pull hf.co/unsloth/Qwen3.6-27B-GGUF:UD-Q4_K_XL

For non-GGUF (safetensors): write a Modelfile with FROM <path-to-gguf> and ollama create <name> -f Modelfile. Conversion to GGUF needs llama.cpp's convert.py first.

Custom Modelfile (e.g. fix Qwen looping with min_p)

Ollama silently drops repeat_penalty/presence_penalty on Qwen 3.5/3.6 (Go runner has no penalty sampler). min_p is the one knob that does work:

ssh <MAC_HOST> '
ollama show qwen3.6:35b-a3b-coding-nvfp4 --modelfile > /tmp/qmf
echo "PARAMETER min_p 0.2" >> /tmp/qmf
ollama create qwen3.6-coder-fixed -f /tmp/qmf
'

Known Mac-specific gotchas

Symptom	Root cause	Fix
ssh <MAC_HOST> ollama → "command not found"	Non-interactive SSH $PATH excludes /usr/local/bin on macOS	Recreate /opt/homebrew/bin/ollama symlink (above)
ssh <MAC_HOST> /opt/homebrew/bin/ollama fails	Old assumption — that path was never canonical	Use /usr/local/bin/ollama or the symlink
Qwen3.5/3.6 tool calls loop or hang	Ollama Go runner has wrong tool-call renderer (Hermes JSON instead of Qwen XML) + drops penalty params + leaves <think> tags unclosed	Use compat.thinkingFormat: "qwen" in pi config + set reasoning: false to skip thinking phase entirely; or switch to non-Qwen models like devstral-small-2
Vision broken on coding-fine-tuned variants	Coding post-training degrades the vision tower; nvfp4 quant compounds it	Use qwen3-vl:32b (purpose-built VL) for vision tasks
ollama ps says model loaded but inference hangs	Stale Metal compute state — usually after another model attempted to load	ollama stop <model> then ollama run <model> to reload clean

Coexistence with Docker Model Runner

Mac runs both Ollama (port 11434) and Docker Model Runner (port 12434) simultaneously. They're independent stacks but share the same Metal GPU — only one model can compute at a time across both. See /docker-switch skill for the Docker side.

claude-local (<WORKSTATION_HOST> zsh function at ~/.zshrc:880) routes through LiteLLM proxy at localhost:4000 → Ollama on Mac (NOT Docker, despite the misleading comment in ~/.litellm/start-proxy.sh).

Related skills

• /docker-switch — Docker Model Runner equivalent
• /ollama-pull — wrapper that pulls + auto-registers (note: <PROJECT_NAME> auto-registration is currently stale; only Anthropic models exposed)
• /ollama-ps — quick "what's loaded" check
• /ollama-list — quick model inventory