A newly published QCon presentation outlines an architecture for mobile applications that use language models without making people wait on a blank chat screen. The central recommendation is to separate model reasoning from interface rendering, allowing a backend to choose from predefined components while the phone displays useful partial results.
A Backend-for-Frontend layer can translate model output into a structured, server-driven interface tailored to the mobile client. Streaming lets text or components appear as they become available, and carefully written prompts can select an appropriate interface without asking the model to generate unrestricted application code. That approach gives product teams more control over accessibility, layout and failure states.
The talk also recommends dividing work between cloud and on-device models. Small local models can handle low-latency or privacy-sensitive tasks, while remote systems perform operations that need more compute or current server data. Multimodal inputs such as images still require explicit loading, progress and fallback behavior in the app.
This is practitioner guidance rather than a new software release or performance study. The patterns do not remove network delays or model mistakes; they make those constraints visible and manageable. Teams still need measurements for time to first useful content, total completion, battery use and incorrect UI selection on the devices they support.