Draw Things adds on-device Qwen 3.5 4B and a clearer mode selector on iOS and macOS
The app says it now runs a lightweight local LLM as an interrogator model and splits generation and editing into distinct modes for a cleaner workflow.
By Ryan Merket · Published
Why it matters
On-device AI in creative tools is moving from nice-to-have to table stakes. A lightweight local LLM can cut latency and keep prompts and assets on-device, while a clearer mode split reduces workflow friction for both new and power users. The combo points to a future where mobile and desktop apps feel faster, leak less data, and make complex AI features easier to control.
Draw Things says its iOS and macOS app now ships a lightweight local LLM, Qwen 3.5 4B, used as an interrogator model, alongside a new Mode Selector that separates image generation and editing. The update was announced in a thread on X from @drawthingsapp.
What shipped
The team behind the Draw Things app is positioning Qwen 3.5 4B as a local, on-device component. In the announcement, it is described as an interrogator model. The provided posts do not include a version number or release notes. The draw is clear, though: a small model that can run on-device can cut latency and preserve privacy compared with cloud calls, which matters for creative tools people use on phones and laptops.
Draw Things runs on iOS and macOS, and this change fits a broader push toward more of the AI stack running locally on Apple hardware. Absent specifics from the post, it is not yet clear whether the LLM operates fully offline by default or whether additional downloads or settings are required.
A clearer workflow
Alongside the model news, Draw Things introduced a Mode Selector that makes the app’s two core activities explicit. "Previously, you were technically always in 'Auto' mode... Now, Draw Things makes the workflow much clearer by separating image generation and editing into distinct modes," the account wrote in the thread on X. One of those is labeled Generation Mode, with editing as a separate path.
For power users, that separation can reduce accidental context switches and make it easier to reason about what the app will do next. For newcomers, it lowers confusion by matching the UI to the mental model: either start from scratch or modify what you have.
The bigger picture
On-device models are becoming a default expectation in creative apps. Even a compact LLM can help with tasks that bridge language and imagery, while keeping prompts and assets local to the device. That makes sense for Draw Things users who want fast iteration without sending data to a server.
There are still open questions. The post does not specify what the interrogator model powers in Draw Things, whether the local LLM is optional or always-on, or any hardware requirements across iPhones, iPads, and Macs. But the direction is consistent: more capability on-device, and a UX that makes the app’s modes explicit instead of implicit.
For builders, the takeaway is straightforward: ship the features users already try to infer, and push intelligence closer to the metal when the workload fits. Draw Things is doing both here, with a small LLM in the stack and a UI that tells you exactly what you are doing.