Running Local LLMs: A Practical Guide to Qwen 3.5-9B on MacBook Pro
Configure your local LLM (e.g., Qwen 3.5‑9B) with thinking mode and tool support to use it as a research assistant without internet.
Set up LM Studio with Qwen 3.5‑9B and enable thinking mode to start using it for code assistance.
Summary
Local LLMs can run on a 24GB MacBook Pro, with Qwen 3.5‑9B (4b quant) achieving about 40 tokens per second when thinking is enabled and a 128K context window. The model requires enable_thinking=true in the prompt template and can be configured in LM Studio or Pi via provider JSON files. Local models lack internet, reducing dependence on big tech, but they also have limited tool use and may loop or misinterpret prompts. The guide details how to set up LM Studio with Qwen 3.5‑9B, including the provider JSON and enabling thinking mode, and how to use OpenCode with the same model.
The article emphasizes that local models are a sustainable alternative, requiring only electricity and a capable machine. It also notes that local models need manual tuning of temperature, top_p, and other parameters to achieve desired behavior. While they may not match the performance of SOTA cloud models, they can serve as a research assistant and code helper.
Overall, the guide provides a step‑by‑step setup for developers wanting to experiment with local LLMs without internet connectivity.
Key changes
- Qwen 3.5‑9B (4b quant) runs on a 24GB MacBook Pro with ~40 tokens/s.
- Thinking mode requires enable_thinking=true in the prompt template.
- LM Studio and Pi can be configured via provider JSON files.
- Local models lack internet, reducing dependence on big tech.
- Tool use is limited; models may loop or misinterpret prompts.
- Local models need manual tuning of temperature, top_p, etc.