What Apex shows about React Native AI models
Featuring
General coding models keep getting better, but React Native still exposes weak spots. Building a static screen is one thing. Working through navigation, animations, async state, native APIs, and performance on devices is another. In this session, Kewin Wereszczynski and Szymon Chmal looked at whether a model built for React Native can hold up against larger frontier models on the kind of work mobile teams deal with every day.
Where Apex differs
Apex grew out of Callstack's React Native expertise, not broad general-purpose coverage. During the stream, Szymon explained how the team used internal know-how, carefully selected React Native libraries with permissive licenses, and training workflows based on supervised fine-tuning and reinforcement learning. The aim was clear: build a model that behaves better inside React Native workflows and understands the framework with more context.
What the benchmarks showed
The session also covered React Native Evals, Callstack's open-source benchmark for testing how models handle navigation, animation, async state, lists, and React Native APIs. Apex finished close to much larger models while remaining cheaper to run. During the stream, we compared Apex with a larger model. The performance gap was relatively small (83.5% versus 87.8%) while the estimated cost dropped from about $22 to $12, making it an interesting tradeoff to explore.
What stood out in the demos
The demos made the comparison easier to judge. Apex answered quickly, handled React Native-specific questions well, and showed why specialization starts to matter once the work moves past static layouts. The conversation also stayed honest about the limits. React Native is hard for models because the UI is only one part of the job. Native behavior, library-level detail, and production constraints still call for stronger context, better evaluation, and solid engineering judgment.
