Local Code by Draw Things, open-source frontier intelligence locally on your Mac

From coding and media creation to documents and everyday tasks, a capable AI agent with everything built in.

We talked about our new app, Local Code, a few weeks ago when we launched its public beta. we’re releasing it to the world and making the entire app open source. You can download it from getlocalcode.app (with a macOS App Store release coming soon), download models from the built-in selector, and start using it today. Rather than reiterate what we’ve already written, I want to talk about what it can do and where it stands—all without external installations, sign-ins, API keys, or external MCP servers.

You can use it to generate images, videos, and sophisticated infographics with precise pie charts and bar charts, using built-in Python and draw-things-cli, entirely locally.

You can use it to create slides and explainer videos with rich motion graphics and animations.

You can use it to write Python or JavaScript, or simply review your code.

You can ask it to visit a website, log in with your credentials, and make an appointment.

You can connect it to your Photos, Messages, Mail, Notes, and Calendar apps, ask it to summarize their contents, or edit and reorganize them where supported.

You can ask it to update your spreadsheets, read Word documents, or make the latest required changes to a PowerPoint file.

You can trust it with your data because it runs the smartest and largest model your Mac can handle today, locally, without sending your prompts or files to a cloud AI provider. DeepSeek 4.1 Flash 1017 is a 763B-parameter model that we quantized into a 376 GiB package. Its frontier-level intelligence surpasses GPT-5.6 Sol on many benchmarks and in real-world use cases. It can run at usable speeds with as little as 36 GiB of resident memory (~350 tok/s for prompt processing and ~18 tok/s for token generation), making it suitable for Macs with 64 GiB or more of RAM. DeepSeek 4 Flash 0731 is a 304B-parameter text-only model that we quantized into an 89 GiB package, suitable for Macs with 48 GiB or more of RAM. Qwen 3.8 27B is a 27B-parameter model that we quantized into a 10 GiB package, suitable for Macs with 16 GiB or more of RAM.

We ran the same benchmarks for these models to compare them with the officially reported results. They closely match those results, with no regression in intelligence.

Benchmark scores for our 2-bit S quantization of DeepSeek 4.1 Flash 1017, demonstrating strong intelligence retention after quantization.

Our quantization format takes full advantage of the M5 Ultra’s Neural Accelerators, delivering leading prompt-processing performance and reducing time to first token compared with alternatives.

Local Code is fully open source. You—or Local Code itself—can clone the repository and modify it to your liking. It’s yours. You can start building Local Code today at https://github.com/drawthingsai/local-code-community.