model-switch

name	model-switch
description	Use when switching the loaded Ollama model on <MAC_HOST> (:11434), checking what's loaded, listing available models, or setting PI's default model. Replaces the docker-switch workflow now that Ollama is the primary local-inference backend. Trigger phrases — "/model-switch", "switch ollama model", "load qwen", "load nemotron", "what's loaded on <MAC_HOST>", "set pi default to".
category	execution
runtimes	["claude"]
pii_safe	true

Ollama Model Switcher

Switch the loaded Ollama model on Mac. Backed by ~/.local/bin/model-switch — a thin curl wrapper around Mac's Ollama API at $OLLAMA_HOST (defaults to http://<PRIVATE_IP>:11434).

Pair: /ollama for general Ollama ops; /docker-switch for the legacy Docker Model Runner path (rarely used now).

Commands

model-switch <alias|full-name>     # warm a model (load into VRAM, ~10–30s for first load)
model-switch list                  # show currently loaded
model-switch ls                    # show all available models on Mac
model-switch unload <name>         # force-unload (sends keep_alive=0)
model-switch default <alias>       # set PI's defaultModel on the LOCAL machine
model-switch help                  # full reference + alias table

From any machine, the script talks to Mac directly via $OLLAMA_HOST. From Mac, it works the same.

Aliases (curated)

Alias	Full name	Notes
qwen36-a3b	qwen3.6:35b-a3b	PI default — primary coding model
qwen36-coding	qwen3.6:35b-a3b-coding-nvfp4	Coding-tuned variant, faster (FP4)
qwen36	qwen3.6:latest	Latest qwen3.6 (Q4)
qwen36-27b	qwen3.6:27b	Lighter qwen3.6
nemotron	nemotron-cascade-2:latest	Long context (was old PI default)
nemotron-3	nemotron-3-super:latest	123B heavyweight reasoner
qwen3-coder	qwen3-coder:30b	MoE coding
qwen235	qwen3:235b-a22b	The big one (142 GB, will swap)
qwen3-4b	qwen3:4b	Voice orchestrator (always-warm via unloader pin)
gemma4	gemma4:latest	Vision/multimodal
lfm2	lfm2:latest	LFM2 MoE
laguna	laguna-xs.2:latest	Laguna XS

Pass any full Ollama name not in the table and it loads as-is.

Common workflows

"Switch to qwen3.6 for coding"

model-switch qwen36-a3b

"What's actually loaded right now?"

model-switch list

"Make qwen3.6 the PI default" (local machine only — run on each box)

model-switch default qwen36-a3b

"Free Mac's GPU"

model-switch unload nemotron      # specific model
# or
ssh <MAC_HOST> ollama stop --all          # all models (use the /ollama skill)

Interaction with the unloader

<LOCAL_PATH>/dev/scripts/ollama-unloader.py (running as ollama-unloader.service on <INFERENCE_HOST>) polls <PRIVATE_IP>:11434/api/ps every 10s. When >1 model is loaded, it evicts non-pinned ones via keep_alive=0. Pinned (PINNED_PAIRS) currently:

• qwen3:4b — voice orchestrator
• qwen3.5:35b-a3b — sub-agent brain
• qwen3.6:35b-a3b — /pdo coding model (added 2026-05-05)

Switching to a non-pinned model and switching away will let the unloader free its VRAM as expected. Pinned models won't get evicted by the unloader — but Ollama's natural 5-min idle keep_alive still applies (so they unload if unused, just on Ollama's clock, not the unloader's).

When to use this vs /ollama

Task	Use
Switch what's loaded	/model-switch
List loaded / available	/model-switch list or ls
Set PI default	/model-switch default
Pull a new model	/ollama (uses ollama pull)
Inspect template / capabilities / params	/ollama
Recreate the /opt/homebrew/bin/ollama symlink	/ollama

Notes

• Requires ~/.local/bin/model-switch on PATH. Distributed by copying to /tmp/model-switch on the target then mv /tmp/model-switch ~/.local/bin/.
• First load of a 30+ GB model takes ~15s (cold) or ~2s (already in OS page cache). Switching between previously-warmed models is sub-second.
• Loading a second model doesn't crash anything on Ollama (unlike Docker Model Runner) — both share VRAM until eviction.