Ollama and opencode setup for the Mac mini

The Ollama LaunchAgent, derived models (mostly to set num_ctx, which
Ollama's OpenAI-compatible API won't take per request), and opencode's
config, now defaulting to qwen3-coder:30b: qwen3:32b tends to end its
turn after thinking without making the tool call it planned.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

README.md

@@ -0,0 +1,42 @@
+# local-llm
+
+Local LLM setup on the Mac mini: [Ollama](https://ollama.com) serving models,
+[opencode](https://opencode.ai) using them.
+
+```sh
+./setup.sh
+```
+
+does everything, and is safe to re-run:
+
+- installs `ollama/com.ollama.server.plist` as the LaunchAgent that runs
+  `ollama serve` (reloaded only when it changed, since that restarts Ollama);
+- builds each model in `ollama/models/` (pulling its base model if missing);
+- symlinks `~/.config/opencode/opencode.json` to `opencode/opencode.json`, so
+  edits made either way end up here.
+
+Needs `brew install ollama opencode`.
+
+## Models
+
+| Model | From | Used for |
+|-------|------|----------|
+| `qwen3-coder:30b-32k` | `qwen3-coder:30b` | opencode's default |
+| `qwen3:32b-32k` | `qwen3:32b` | opencode, as an option |
+| `inspector` | `qwen3:32b` | 16k-context variant |
+
+Each derived model mostly exists to set `num_ctx`. Ollama's OpenAI-compatible
+API, which opencode talks to, ignores any context size sent with a request, so
+without it a long agent prompt is silently cut to Ollama's small default
+context. A Modelfile's `# name:` line is the name it's created under.
+
+`qwen3:32b` is a thinking model, and in opencode it often ends its turn after
+thinking, without making the tool call it planned. `qwen3-coder` was trained
+for agentic tool use and doesn't think first, so it doesn't have that
+problem, and it's much faster: a mixture of experts with ~3B active parameters.
+
+## Server settings
+
+From the LaunchAgent: listens on all interfaces (`0.0.0.0:11434`), keeps models
+loaded indefinitely (`OLLAMA_KEEP_ALIVE=-1`), flash attention on. Logs go to
+`/tmp/ollama.log`.

ollama/com.ollama.server.plist

@@ -0,0 +1,14 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0"><dict>
+  <key>Label</key><string>com.ollama.server</string>
+  <key>ProgramArguments</key><array><string>/opt/homebrew/bin/ollama</string><string>serve</string></array>
+  <key>EnvironmentVariables</key><dict>
+    <key>OLLAMA_HOST</key><string>0.0.0.0:11434</string>
+    <key>OLLAMA_KEEP_ALIVE</key><string>-1</string>
+    <key>OLLAMA_FLASH_ATTENTION</key><string>1</string>
+  </dict>
+  <key>RunAtLoad</key><true/>
+  <key>KeepAlive</key><true/>
+  <key>StandardErrorPath</key><string>/tmp/ollama.log</string>
+</dict></plist>

ollama/models/inspector.Modelfile

@@ -0,0 +1,3 @@
+# name: inspector
+FROM qwen3:32b
+PARAMETER num_ctx 16384

ollama/models/qwen3-32b-32k.Modelfile

@@ -0,0 +1,5 @@
+# name: qwen3:32b-32k
+# General-purpose thinking model. In opencode it often ends its turn after
+# thinking without making the tool call it planned; fine for plain chat.
+FROM qwen3:32b
+PARAMETER num_ctx 32768

ollama/models/qwen3-coder-30b-32k.Modelfile

@@ -0,0 +1,5 @@
+# name: qwen3-coder:30b-32k
+# opencode's default: trained for agentic tool calling, no thinking phase,
+# mixture-of-experts (~3B active params) so it's fast.
+FROM qwen3-coder:30b
+PARAMETER num_ctx 32768

opencode/opencode.json

@@ -0,0 +1,26 @@
+{
+  "$schema": "https://opencode.ai/config.json",
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama (local)",
+      "options": {
+        "baseURL": "http://localhost:11434/v1"
+      },
+      "models": {
+        "qwen3-coder:30b-32k": {
+          "name": "Qwen3 Coder 30B (32k context)",
+          "tools": true,
+          "limit": { "context": 32768, "output": 8192 }
+        },
+        "qwen3:32b-32k": {
+          "name": "Qwen3 32B (32k context)",
+          "tools": true,
+          "limit": { "context": 32768, "output": 8192 }
+        }
+      }
+    }
+  },
+  "model": "ollama/qwen3-coder:30b-32k",
+  "small_model": "ollama/qwen3-coder:30b-32k"
+}

setup.sh

@@ -0,0 +1,38 @@
+#!/bin/sh
+# Makes this Mac match the repo: Ollama's LaunchAgent, the derived models in
+# ollama/models, and opencode's config (symlinked, so edits land in the repo).
+# Safe to re-run.
+set -eu
+cd "$(dirname "$0")"
+repo=$(pwd)
+
+# Ollama server LaunchAgent. Only reloaded when it changed: reloading
+# restarts Ollama and unloads whatever models are in memory.
+plist=~/Library/LaunchAgents/com.ollama.server.plist
+if ! cmp -s ollama/com.ollama.server.plist "$plist"; then
+  echo "Installing $plist"
+  launchctl bootout "gui/$(id -u)" "$plist" 2>/dev/null || true
+  cp ollama/com.ollama.server.plist "$plist"
+  launchctl bootstrap "gui/$(id -u)" "$plist"
+  until curl -fs localhost:11434/api/version >/dev/null; do sleep 1; done
+fi
+
+# Derived models: each Modelfile names its model on a "# name:" line.
+for f in ollama/models/*.Modelfile; do
+  name=$(sed -n 's/^# name: *//p' "$f")
+  base=$(sed -n 's/^FROM *//p' "$f")
+  [ -n "$name" ] && [ -n "$base" ] || { echo "$f: needs '# name:' and FROM lines" >&2; exit 1; }
+  ollama show "$base" >/dev/null 2>&1 || ollama pull "$base"
+  echo "Creating $name from $f"
+  ollama create "$name" -f "$f" >/dev/null
+done
+
+# opencode
+mkdir -p ~/.config/opencode
+target=~/.config/opencode/opencode.json
+if [ -e "$target" ] && [ ! -L "$target" ]; then
+  mv "$target" "$target.bak"
+  echo "Moved the old $target to $target.bak"
+fi
+ln -sfn "$repo/opencode/opencode.json" "$target"
+echo "Linked $target"