A 0.5B model on this Mac writes Python from plain-English requests and the code runs: Qwen2.5-Coder-0.5B-Instruct on Core ML (qwen2.5-coder-0.5b-coreml), the prompt on the Neural Engine and the writing on the GPU. Nothing leaves the machine.
swift run -c release CodeWriterDemo # downloads the model (~1 GB) on first launch
Sources/CodeWriterDemo/demo.sh --autostart # same app, plus one terminal: macmon above the live log
CODE_WRITER_MODEL_DIR=<folder> swift run -c release CodeWriterDemo # a local copy instead of the downloadPlay (⌘↩) works down a list of ten short tasks, one per Python feature (built-ins, list comprehension, slicing,
dict, string methods, conversion, tuples, loops, sorted, all()). Each one is written live into the editor, then
python3 runs the task's three asserts and the console shows each result. Pause stops after the current task;
Replay (⌘R) starts over. Type your own request in the bar at the bottom: it is written the same way and checked
to parse as Python (it has no tests). Running the model's code needs python3 on the PATH.
The tasks come from MBPP (sanitized test split, CC-BY-4.0), with the first assert in the prompt so the model knows the function name. The list is hand-picked for the recording: one original pick (sort a matrix by row sum) failed and was replaced with one that passes. All ten pass; the unbiased numbers are below.
M5 Pro, macOS 27, greedy decoding:
| The ten demo tasks | 10/10 pass all asserts, ~50 s in total (95–330 tokens each) |
| Prompt (448 tokens, Neural Engine) | ~52 ms |
| Writing (GPU) | 36–51 tokens/s |
| MBPP sanitized test split (257) | 119/257 = 46.3% |
| HumanEval (164) | 89/164 = 54.3% (PyTorch fp32: 90/164 = 54.9%) |
CodeWriterCheck replays HumanEval through the Swift host and compares every token stream with a reference run.