Qwen3.5-2B Runs Natively on M1 Pro

This is the kind of post that actually matters for indie AI builders: it turns a capable small model into a runnable local workflow on consumer Mac hardware. The caveat is that the setup details need care, because the difference between MPS and CPU fallback is the difference between a usable demo and a slow toy.

• –Qwen3.5-2B is a 2B-parameter checkpoint, so it fits the “small enough to iterate locally” niche the Qwen model card targets for prototyping and development.
• –For Apple Silicon users, the real value is forcing Metal acceleration; without that, this kind of setup quietly degrades into CPU inference and loses the point.
• –Wrapping the model in Gradio makes it immediately useful as a local sandbox for prompt tests, tool prototyping, or lightweight internal apps.
• –The post is less about a novel model breakthrough and more about lowering the friction to use open-weight models in everyday Mac dev environments.