Serious AI capability inside your walls

How does a banking assistant turn a customer’s question into an authorized action? Follow an unfamiliar payment from transaction lookup to card freeze and specialist handoff, exploring how models, banking services, and human teams work together and what the bank needs to control at each step.

Apex is now publicly available. Built by Callstack, it brings React Native expertise to feature development, code reviews, debugging, and testing, with strong performance on Next.js tasks too. See the benchmark results, compare API pricing, explore self-hosting options, and get started with Apex in your own project.

Apple's 2026 AI updates move Apple Intelligence beyond a narrow text model. Foundation Models gets image input and server-backed calls, high-end devices get a stronger on-device model, and Core AI gives custom local models a native path. We're mapping these APIs into the React Native AI stack and working with early users on real app workflows.

We optimized self-hosted Gemma 4 31B for Apex by tuning vLLM server configuration, GPU memory use, KV cache capacity, context length, multimodal limits, tensor parallelism, and MTP speculative decoding. The article covers what changed, the performance gains we measured, the latency tradeoffs, and the production setup we recommend for faster Gemma serving.

Today, we are introducing Apex, our first domain-specific coding model designed for React Native. Built on Gemma 4, Apex achieves frontier-level performance at a fraction of the cost of generalized models. It is optimized for the fast, agentic workflows our teams run daily, delivering up-to-date results with fewer tool calls.

Apple Foundation Models are most useful on mobile when the task is small, fast, local, and not worth a cloud round trip. The real value is not replacing cloud models entirely, but using on-device AI where latency, privacy, and frequency matter most.

Not every LLM system needs a full RAG stack. This article shows when structured context injection can replace retrieval pipelines, reducing complexity, latency, and operational overhead.

The same LLM can appear in many variants because different hardware environments require different numerical precision. This article breaks down quantization, explains model naming conventions, and shows how deployment constraints shape model formats. Explore how quantization impacts performance and infrastructure costs.

React Native Evals is a new open-source benchmark from Callstack that evaluates how coding models implement real React Native tasks. The suite ships with 39 evals across animation, async state, and navigation categories, covering libraries like Reanimated, TanStack Query, Zustand, and React Navigation.

MLC-LLM can run LLMs fast on-device, but we hit a nasty Android issue: the first inference froze the system UI for up to 50 seconds, then everything was fine. This post walks through how we added Perfetto-compatible traces in MLC-LLM and kernel profiling in TVM/OpenCL, and why switching from _1 to _0 model formats fixed it on Adreno.

We’ve added on-device speech synthesis to our Apple provider for the AI SDK, completing the feature set alongside generation, embeddings, and transcription. This integration with AVSpeechSynthesizer brings fast, private, and offline text-to-speech to React Native, with support for multiple voices, languages, and even Personal Voice on iOS 17+.

This post introduces speech transcription support in React Native AI, implemented using Apple’s SpeechAnalyzer and SpeechTranscriber APIs available in iOS 26. It explains how to use the AI SDK’s transcribe function to pass audio buffers, how modular analysis is configured through the SpeechAnalyzer class, and how system-level asset management ensures models are handled outside the app bundle.

Our React Native AI library now supports on-device text embeddings by integrating Apple's native Natural Language framework. This implementation leverages the NLContextualEmbedding model built directly into iOS. Because the model is treated as a shared operating system asset, it enables powerful semantic features without requiring developers to bundle large files or force users to perform any extra downloads. We expose this capability through our Vercel AI SDK provider, simplifying integration for React Native developers.

We’ve just added tool calling to our Apple provider for the AI SDK, unlocking the ability to run custom functions directly within Apple’s on-device language model. This means your React Native apps can now generate dynamic, interactive content, like fetching live data or executing app-specific logic, without ever leaving the device.

Apple’s on-device LLM is now available in React Native, enabling developers to build private, fast, and offline AI features powered by Foundation Models. This preview release supports text generation, streaming, structured outputs, and first-class support for Vercel AI SDK. A stable version is planned for release alongside iOS 26 later this fall.

Running LLMs locally on mobile isn’t a fantasy-it’s possible with MLC. This post explains how MLC compiles both model and runtime to create optimized, cross-platform packages for local inference. Discover what’s required, how it works, and why MLC stands out among edge ML frameworks.

Running AI models locally in React Native used to sound impossible, but not anymore. This post introduces react-native-ai, a new library powered by MLC and integrated with Vercel AI SDK, enabling on-device LLM inference with a web-like API and native performance.

Running large language models directly on mobile devices is currently limited by model size, format compatibility, and performance trade-offs. However, tools like MLC are emerging to simplify the process and pave the way for a future where everyone has their own AI assistant in their pocket.

Callstack shares lessons learned from building a video transcription app using OpenAI's Whisper for speech recognition. The guide covers extracting audio, generating transcripts, speaker diarization, and testing different Whisper model sizes to optimize for speed and accuracy.

Learn how React Native AI uses a unified provider model to switch between cloud and on-device LLMs without rewriting application logic, enabling reliable AI experiences online and offline.

Learn how on-device models in React Native can call tools, interact with external APIs, and return structured outputs you can use directly in application logic.

Learn how to run third-party on-device LLMs in React Native using MLC LLM. Choose your models, run them efficiently across platforms, and keep a unified JavaScript API with React Native AI.

Learn how to use Apple’s built-in Foundation Models in React Native apps, reduce memory usage, and enable fast, fully on-device AI with React Native AI Apple.

Learn why on-device AI matters for React Native apps, how local LLMs behave offline, and what problems React Native AI is designed to solve from day one.

Artur Morys-Magiera and Lech Kalinowski share findings from 344 benchmark runs with Apex, our Gemma-based LLM, examining the trade-offs between model size, decode speed, and quality.

Explore Jev with Kewin Wereszczyński and Szymon Chmal: how it works, its approach to faster agent decisions, and possible local alternatives.

Lech Kalinowski will show how Agents Commander lets multiple AI coding agents collaborate directly in one terminal without a central server or heavyweight framework.

See how Apex, Callstack's React Native model, compares with frontier models on navigation, animations, async state, and performance-heavy UI work.

Join Lech Kalinowski at AI Engineer Miami on April 20–21 for a talk on running offline latent diffusion models on mobile NPUs.

React Native Evals introduces a data-driven way to measure how AI coding models perform on real React Native development tasks.

A practical look at reliability, performance, libraries, and tradeoffs when running LLM inference locally in React Native apps.

Learn how to run on-device LLMs in React Native using Vercel’s AI SDK from Michał Pierzchała's talk at DevAI by Data Science Summit.

In this talk, Szymon will show how to run LLMs directly inside React Native apps using an AI SDK that provides a powerful abstraction layer to simplify building AI applications. Join him as he explores react-native-ai, a library that enables local LLM execution.

Need offline AI in your mobile app? This talk recording demonstrates how to use the AI SDK to create a custom provider for on-device LLMs in React Native, ensuring privacy and functionality anywhere.

On Sept 2 in Wrocław, join top speakers from around the world at the AI Meetup. Learn about AI SDK 5, on-device LLMs in React Native, TV-ready apps with Kiro.dev, and how AI supports developers beyond code.

Tune in to dev chat on Apple’s AI SDK, iOS 26 naming chaos, Siri’s limits, Liquid Glass UI, and on-device LLMs.

Watch how MLC brings on-device AI to React Native. See react-native-ai in action and discover what’s next for local model execution.

Leveraging capabilities of local devices to run LLMs via React Native and Vercel AI SDK.