The team behind the Draw Things image app has taken Local Code out of beta and open-sourced it. It is an AI agent app for Mac that runs a large model on the machine, and the team says it works without external installs, sign-ins, API keys or external MCP servers. The download is at getlocalcode.app, with a Mac App Store release “coming soon”. The source is GPL-3.0 and also builds for iPhone and iPad.
The models
Local Code downloads its own quantized checkpoints and suggests one by memory: Qwen 3.8 27B below 48 GB (2-bit under 32 GB, 4-bit from 32 GB), DeepSeek 4 Flash 0731 from 48 GB, and DeepSeek 4.1 Flash 1017 from 64 GB. The 1017 build is a 376 GB download in 2-bit form, which includes a 203 GB engram file. The app keeps the most-used experts in memory and streams the rest from disk, capped at the Mac’s memory minus 32 GB. The README says you can also pick a cloud model from Anthropic, OpenAI, OpenRouter or an OpenAI-compatible server, and that your prompts and files then go to that provider.
On an M5 Max with 128 GB, the README’s table shows DeepSeek 4.1 Flash 1017 at about 346 tokens a second of prefill and 18.7 tokens a second of output with 8K tokens already in context. On an M5 Ultra with 256 GB it reports about 4,600 and 58.9. These are Draw Things’ own runs.
The sandbox
The shell, Python 3.14, Node.js 24, LibreOffice, FFmpeg and the Draw Things CLI for images and video are compiled into the app and run inside Apple’s App Sandbox. The agent works in a workspace folder you pick, has no full-disk access and cannot launch other programs. Network commands such as web search, curl and ssh still connect out. Built-in commands read Calendar and Photos and can edit them. On a Mac, notes, mail and imessage can read Notes, Mail and Messages but, per the README, “can’t change or send anything”, which is narrower than the launch post’s wording. A WebKit browser the agent drives can use sites that need a sign-in.
The quality claim
Draw Things says its quantized models closely match officially reported results. The README’s table checks three benchmarks, AIME 2026, MathArena Apex and GPQA Diamond, and compares its 1017 build with the scores DeepSeek reported for an earlier checkpoint, 4.1 Flash 0910. That is a useful sanity check, not a full evaluation.
Why this matters
If you have an Apple Silicon Mac with 64 GB or more, this is one of the few agent apps that bundles a large open model, a sandboxed toolchain and media generation with nothing sent to a cloud provider by default. Check the folder and personal-data permissions you grant before you point it at real mail or photos. Owners of non-Apple machines get nothing to run yet, since the engine targets Apple Silicon. For the model itself, see our coverage of DeepSeek’s V4.1-Flash-1017 weights and an independent safety test of the same build, which matters more once the model has a shell.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment