Local Code: Draw Things brings an open-source AI agent to the Mac, running locally

What's going on

Draw Things, known for an app for local AI image generation on Apple devices, has released "Local Code": an open-source AI agent for the Mac. An agent is an AI that doesn't just answer questions but carries out tasks on its own, such as writing programs, editing files or acting in a browser. The release post is dated October 28, 2026, and MacRumors reported on it on October 30. Local Code runs Qwen and DeepSeek models on your own Mac. A Mac App Store version is due soon.

What the agent can do

According to the release post, models are downloaded through a built-in picker. You don't need external installs, sign-ups, API keys or extra MCP servers (interfaces that let agents connect to tools). Tasks include:

• Generating images, videos and infographics, using built-in Python and the draw-things-cli tool
• Creating slides and explainer videos
• Writing or reviewing Python and JavaScript code
• Updating spreadsheets, Word and PowerPoint files
• Using a built-in browser, for example to log in to a website and book an appointment
• Working with Photos, Messages, Mail, Notes and Calendar to summarize content or, where possible, change it

MacRumors adds: Mail, Messages and Notes are only read with the user's permission, and calendar entries are only changed with permission. According to the documentation, Apple's App Sandbox (a protection mechanism that isolates apps from the rest of the system) restricts access to files and system areas.

Models and hardware

Quantized means the model's numbers are stored at lower precision. That saves space and can cost some quality. According to Draw Things, the largest model can run with only 36 GiB of memory actually in use. It reaches about 350 tokens per second when reading the prompt and about 18 tokens per second when writing (a token is a chunk of text, roughly a syllable). The package itself still needs 376 GiB of disk space. MacRumors cites downloads of around 89 GB or 376 GB for the large DeepSeek models and 10 to 21 GB for the recommended smaller ones.

Vendor claims about quality

The developers say DeepSeek 4.1 Flash 1017 beats GPT-5.6 Sol on many benchmarks and in everyday use. They say their own reruns of the tests closely matched the official results, and that quantization causes no loss of intelligence. The figures shown are for a 2-bit S quantization. They also say the format uses the Neural Accelerators of the M5 Ultra and offers leading prompt processing and a shorter time to first token. These are all self-reported claims, and no independent measurements are cited.

Privacy

According to MacRumors, with local models your prompts and files are processed on the Mac and not sent to a cloud provider. If you pick a cloud model, prompts and relevant files go to the provider you choose. So the benefit depends on which model you select. If you handle sensitive data, make sure a local model is active. Because the code is open, you can inspect how data is handled.

What you need to try it

• A Mac with an Apple silicon chip. A Windows/Linux PC or Google Colab won't work, since the app is built for the Mac.
• For Qwen 3.8 27B: at least 16 GiB of RAM and about 10 GB of free storage.
• For DeepSeek 4 Flash 0731: 48 GiB of RAM or more and around 89 GB of space.
• For the largest model: 64 GiB of RAM or more and 376 GiB of space.
• A fast internet connection for the downloads.

Our take

A fully local, open agent with a browser, Office files and system access is appealing for Mac users who want to avoid cloud services and subscriptions. Getting started with the 27B model is realistic on many Macs. Be cautious about the claims of a lead over GPT-5.6 Sol and of lossless quantization: neither has been independently confirmed. Tests will also have to show whether 18 tokens per second is practical, and how safe an agent that edits your mail and calendar really is. The models themselves come from Qwen and DeepSeek. Draw Things supplies the app.