Running a real agent on Qwen 3.5 9B on an M1 Pro MacBook (16GB)

A personal automation system built on a Claude Code-style agent is deliberately model-agnostic, so switching it to a local model was a one-line config change. The author pointed it at Qwen 3.5 9B through Ollama on an M1 Pro MacBook with 16GB of unified memory, and ran real tasks from their own queue rather than chat demos. Ollama exposes an OpenAI-compatible API on localhost:11434, so nothing else changed.

What worked, and what did not

• Memory recall: reading structured memory files, finding the relevant part and reporting it. A 9B model is genuinely fine for this.
• Tool calling: reasonable on straightforward requests; it invoked the right tools most of the time. The author notes this matters more than text quality for automation.
• Creative and complex reasoning: a noticeable gap, as expected. The point is not beating a frontier model; it is whether a real subset of agent work can run without a cloud API, and it can.
• Speed was "within acceptable range": noticeable, but not a blocker (no token-per-second figure was published).

Bonus: on-device phone

The author also ran Qwen 0.8B and 2B on an iPhone 17 Pro with PocketPal AI (open source), downloading the model once and then enabling airplane mode; it still answered, with nothing leaving the device.

The framing is the interesting part: not every agent task needs a frontier model. Reading a file, formatting output, summarizing a short note or routing a request can run locally without per-token cost or sending personal data anywhere.