Qwen Coder and Devstral are strong starting points for local coding on Apple silicon. The right model depends on whether you need completion, repository questions, or an agent that edits files and runs tests.
Picks by unified memory
| Mac memory | Candidates | Best use |
|---|---|---|
| 16 GB | small Qwen Coder or 7B FIM model | completion and small functions |
| 24 GB | quantised Devstral Small 2 | bounded agent runs |
| 32–64 GB | Qwen3-Coder 30B-A3B, Devstral Small 2 | repository work, tests, refactoring |
| 96 GB and above | Qwen3-Coder Next and larger variants | larger context and demanding agents |
Qwen3-Coder 30B-A3B activates only part of its mixture-of-experts weights per token, which can make it much faster on a Mac than a dense model of similar total size. Devstral Small targets agentic software engineering but reads more weights for each generated token.
Completion needs fill-in-the-middle support. Agents need reliable tool calls, patches, and tests. We compare candidates on identical tickets, tools, and budgets, measuring accepted diffs, passing tests, denied file access, latency, and review effort.
Read local AI coding on Mac and M6 versus M5 for local AI.
Compare coding models on your repository
We test candidates on real tickets, select the Mac, and integrate a local endpoint with bounded permissions.