Raph Levien on Rust GUI: AI drives interest, GPUI and native mobile
Raph Levien has published "Advice for the next 100 Rust GUIs" on linebender.org, dated November 21, 2026. It follows his 2022 post "Advice for the next dozen Rust GUIs", published 4.5 years earlier. He says roughly 20 Rust GUI toolkits have been announced since then, and that interest in Rust GUI has recently accelerated, fueled by AI.
Why Rust GUI is still compelling: many shipping GUI apps perform poorly, especially cross-platform ones. Platform-shipped toolkits have largely stagnated, and system vendors are not interested in the cross-platform problem, so Electron has grown. With RAM still expensive and focus shifting to mobile, bloat has real consequences. Adoption of Rust UI has nonetheless been slow, and few flagship apps use it. One of the most notable is the Zed editor, built on its own GPUI framework. Levien says GPUI is emerging as a major choice for AI-driven rewrites, most notably Discord. His personal sense is that GPUI suits the Zed use case, and that shipping in a real project boosts its credibility, but it is not necessarily ideal for more general purpose apps. Its 2D renderer suits a text editor but lacks full gradient support, blending, filters and other features expected of a renderer in the same class as Vello or Skia.
He argues the causes of slow adoption may be fading. Rust is harder to write than TypeScript, Kotlin or Swift, and engineers skilled in both Rust and UI are scarce. But he says there is evidence that the LLM cost of writing Rust is not significantly higher than for other languages, and reason to believe Rust will be more maintainable in large codebases because its type system helps enforce invariants. If costs are about equal, why not pick the language with a much better mix of safety and performance? In his words, Rust is rapidly becoming the language for AI rewrites of existing codebases, yet there is no consensus on the UI layer. He predicts a small explosion of trying lots of different things.
On "native": his 2022 point that the term has no clear meaning still largely holds, but it depends on the platform. Windows has no clear native choice. On the Mac there is some fragmentation between AppKit and SwiftUI, and he cites a 2025 Michael Tsai blog post on SwiftUI performance issues. Mobile differs, because platforms keep investing and fine details matter. Liquid Glass on iOS comes with little effort in UIKit or SwiftUI, while cross-platform apps won't quite seem right. Making gesture recognition, scroll dynamics and IME 100% match the platform is not yet fully solved. So for mobile he calls an extremely viable path native UI with common business logic in Rust. He says this is what Shopify recently chose, abandoning React Native (he is unsure about the business logic part), that he believes 1Password has this architecture, and that it is Discord's "Phoenix" plan for mobile, with GPUI on desktop. There has also been considerable progress on building mobile UI entirely in Rust: the editor demo in android-view has working IME features (swiping, moving the cursor by rolling the space bar, deleting in semantic chunks), composition regions needed for CJK, and basic accessibility through AccessKit.
Infrastructure: text layout, an unsolved problem in 2022, now has a good solution in Parley. It got there through sustained investment from Google, Canva and Dioxus (now Cognition), plus community contributions, and seems to be used in the majority of new Rust UI toolkits. Cosmic Text is a viable alternative that reduces monoculture risk. AccessKit is now solid infrastructure used by most toolkits, though he distinguishes good integration from box ticking, and says rich text editing is not yet fully accessible. He also wants the performance drop with accessibility enabled fixed. For egui it is a particular problem, since it has to build and diff a tree every animation frame, which undercuts its immediate mode design.
The system compositor is still unsolved, and more pressing given the importance of video: a video should play in a hardware accelerated layer where possible, and on low end devices that is the only hope of smooth frame rates. The subduction library is notable progress but is not commonly integrated into toolkits. He sketches a rich, high level interface like CALayer on Apple platforms, with polyfills elsewhere; on Android, simple overlays would use SurfaceFlinger layers and more sophisticated compositing would run shaders on the GPU. He wants it built as a high quality community project rather than dozens of vibe-coded forks, though even that beats nothing.
Immediate versus retained mode: he suggests one general purpose UI for everything may be behind us, especially as LLMs can tune for specific use cases. Immediate mode is appealing because it needs less code and seems simpler, and it works well for things like game development tooling. But he is concerned it won't scale: it is good at the first 90% but makes the second 90% harder. To be production-ready it needs stable widget identities for the accessibility tree and reliable caching of expensive artifacts such as text layouts, and it pulls toward re-rendering everything each frame instead of using compositor layers. For these reasons he is not impressed that the ArtCraft projects chose egui. He still calls egui a really good immediate mode implementation that solves several problems, including accessibility, which mainline Iced does not yet support, but says it would not be his first choice for productivity and design apps.
Modules: LLM-assisted engineering works best with clearly scoped modules with solid specifications, where an LLM given clear objectives such as performance and robustness can do very well. He wants a future separation of design and execution, with a good spec between them, and recommends adapting web standards without following them slavishly. Web standards bring care, cross-platform work and the extensive WPT tests, which helps explain Electron's popularity. But DOM is slow and memory-intensive, so a lighter tree with a bulk mutation interface might serve GUI apps better. Full CSS resolution and layout are expensive, so a curated subset might help; he has experimented with selectors as an incremental state machine, at the cost of features like sibling selectors. The Web also lacks a solid way to represent virtualized lists. He sees Blitz, a modular browser engine, as well placed to anchor hybrid approaches, for example starting with Stylo and later swapping in a leaner CSS engine. He cites the Xilem/Masonry split as a good example of modularity: Xilem's reactive layer is independent of the elements it controls, with instantiations for the DOM and for Masonry widgets, and Masonry is independent of the reactive layer and the language above it. By contrast, he says SwiftUI mashes together the widget implementation, the reactivity framework and the language.
Key facts
- Raph Levien's November 21, 2026 post follows his 2022 "Advice for the next dozen Rust GUIs"; he counts roughly 20 Rust GUI toolkits announced since, and a recent AI-fueled surge of interest.
- He says GPUI (Zed's framework) is emerging as a major choice for AI-driven rewrites, most notably Discord, but its renderer lacks gradients, blending and filters and may not suit general purpose apps.
- For mobile he calls native UI with shared business logic in Rust an extremely viable path, citing Shopify, and says Discord's "Phoenix" plan takes this route while using GPUI on desktop.
- Text layout (Parley) and accessibility (AccessKit) are described as basically solved infrastructure; the system compositor is still unsolved, with the subduction library as early progress.
- He is concerned immediate mode, including egui, won't scale well, and advises building modular toolkits that adapt web standards rather than follow them slavishly.
Why it matters
Levien's argument is that AI is changing the economics of Rust UI work. He says the LLM cost of writing Rust is not significantly higher than for other languages and that Rust may be more maintainable in large codebases, so Rust is rapidly becoming the language for AI rewrites. The UI layer has no agreed strategy, and he expects a small explosion of experiments. The post also takes stock of the roughly 20 toolkits announced since 2022 and of which hard problems are now solved.
Who it affects
Authors of Rust GUI toolkits, who get opinionated advice on rendering, text, accessibility, compositing and modular design. Teams planning AI-assisted rewrites into Rust, such as the Discord case he describes. Mobile developers deciding between native UI and cross-platform toolkits. Developers choosing between egui, Iced, GPUI and the Xilem and Masonry stack.
How to use it
The post does not recommend a single toolkit. Its practical guidance: on mobile, consider native UI with common business logic in Rust. Use Parley for text layout, with Cosmic Text as an alternative, and AccessKit for accessibility, aiming for real integration rather than box ticking. Be wary of immediate mode for large, production applications. Separate design from execution behind a clear spec, and borrow from web standards selectively, for example Blitz with Stylo that can later be replaced by a leaner CSS engine.
How solid is it
This is an opinion essay published on linebender.org under Levien's byline, not a benchmark or study. Several claims are hedged or uncited: the evidence on LLM cost of writing Rust is not cited, 1Password's architecture is stated as "I believe", and the GPUI fit for Zed is described as his "personal sense". No measured performance numbers are given for egui with accessibility, SwiftUI or DOM. This account covers the essay through its section on modules.
Risks and caveats
Much of the advice is judgment and prediction: the "small explosion" of approaches, the concern that immediate mode won't scale, and the view that GPUI is not necessarily ideal beyond text editors. Levien does not say GPUI is a bad choice in general. The Shopify and 1Password architectures are his reading, not confirmed by the companies in the post. He also warns that a compositor library could end up as dozens of vibe-coded forks rather than one well-built community project, and that accessibility work such as rich text editing is still incomplete.
“it’s good at the first 90%, but it makes the second 90% harder”
— Raph Levien, on immediate mode GUI, linebender.org