Local Code ships free on Mac on October 30, 2026
On October 30, 2026, Draw Things' Local Code ships free on Mac with quantized DeepSeek and Qwen models.
| DeepSeek 4.1 Flash 1017 | DeepSeek 4 Flash 0731 | Qwen 3.8 27B | |
|---|---|---|---|
| parameters | 763B | 304B | 27B |
| package size | 376 GiB | 89 GiB | 10 GiB |
| suitable RAM | 64 GiB or more | 48 GiB or more | 16 GiB or more |
| resident memory | 36 GiB | ||
| prompt processing | ~350 tok/s | ||
| token generation | ~18 tok/s |
Local Code downloads models and runs them on the Mac.
Qwen 3.8 27B starts at 10 GiB and needs a 16 GiB Mac, while larger DeepSeek downloads run to 89 GiB and 376 GiB.
DeepSeek 4 Flash 0731 needs 48 GiB in an 89 GiB package, and the largest DeepSeek wants 64 GiB or more.
The 376 GiB DeepSeek 4.1 Flash 1017 package does not prove that a 64 GiB Mac will run it at a useful token rate.
The release says the 763B-parameter model can run at usable speeds with as little as 36 GiB of resident memory, about 350 tokens per second for prompt processing and about 18 tokens per second for generation.
Local use is free, while optional cloud access uses the user's API keys or prepaid credits starting at $9.99.
A direct Mac download is available now, and the developers say a Mac App Store release is coming soon.
The document does not prove that quantized weights preserve frontier-level intelligence, because it reports no benchmark scores, only that its results closely match official numbers.
Power draw, sustained throughput, App Store timing, and the benchmark deltas behind the 2-bit S quantization claim remain unmeasured.
The document proves package sizes and a memory floor, not that quantized intelligence survives.
After huggingface.co. We did not report this. The pictures, if any, are theirs.
