A software company can run coding assistance on a Mac mini, Mac Studio, or MacBook Pro without sending its repository to an external model API. That matters for client code, unreleased products, and internal automation. It does not mean every local model replaces Claude or another frontier model for difficult architecture work.
Memory guidance for local coding models
- 24 to 32 GB: smaller coding models, completion, tests, and bounded repository questions.
- 48 to 64 GB: roughly 30B-class models, more context, and a small shared endpoint.
- 96 to 128 GB: larger models, several tools, or greater concurrency.
Coding agents carry system instructions, source excerpts, and tool results in context. Leave memory for the KV cache, runtime, IDE, and build processes rather than sizing from model weights alone.
A controlled local stack
Ollama, llama.cpp, or Apple’s MLX serves the model locally. An IDE extension or agent connects only to that endpoint. Give it the minimum tools required: read approved files, propose diffs, and run defined tests.
If local processing is a binding data boundary, disable cloud features and inspect telemetry, crash reports, and package downloads. “Runs on the Mac” is not enough if another component sends the prompt elsewhere.
Current specialist families such as Qwen Coder and Devstral illustrate why a code-tuned model can outperform a similarly sized general chat model on coding tasks. Apple’s MLX is designed around Apple silicon unified memory, while Ollama provides a straightforward local server.
Evaluate models on your repository
Ask at least two candidates to:
- fix a faulty function and its test,
- add a test using existing conventions,
- explain a change spanning several files,
- respect denied secrets and directories,
- run the build reproducibly.
Accepted changes per engineering hour are more useful than tokens per second.
Keep operating knowledge inside the company
Many commercial business plans state that customer content is not used for model training by default. An external service still processes prompts and code under its contractual, retention, and security terms. A strictly local endpoint keeps the repository, corrections, and internal working patterns outside that provider path.
Your company controls model versions, updates, logs, and deletion. It also owns the evaluation set that makes future model changes measurable.
Limits and a sensible hybrid
Local models may be slower or weaker on open-ended architecture questions. One Mac also has finite concurrency. Local agents can still make incorrect edits or propose destructive commands. Use isolated working trees, least privilege, code review, and CI.
A hybrid can keep private routine work local while allowing approved abstract questions with no confidential context to use a cloud model. Our local AI on Mac service covers model evaluation, hardware sizing, and a controlled development integration.
Test local coding AI on your repository
We compare models on representative tasks, size the Mac, and configure a local endpoint with bounded tool permissions.