Web, iOS, Android, React Native, Flutter, and .NET — six SDKs, nine backends, one HTTP surface. Zero cloud cost. Zero install for your users.
Web, iOS (Swift), Android (Kotlin), React Native, Flutter, and .NET (MAUI / Avalonia / WinUI / Catalyst / desktop). Same OpenAI HTTP contract on every platform; switch the import, keep the agent code.
WebLLM, Transformers.js, llama.cpp, Apple Foundation Models, CoreML, MLX, MediaPipe LLM, LiteRT, and (.NET only) ONNX Runtime + ML.NET. Picked automatically per platform; opt into specifics when you need to.
Blank-chunk detection, generation timeout, and a bounded recovery loop that unloads + reloads the engine on fatal errors. Users see a brief reload instead of a broken UI.
React Hooks for the JS family; idiomatic Swift actors, Kotlin coroutines, Pigeon-typed Dart APIs, and async C# Tasks for the native SDKs. No JS-side state machine to learn.
Works in Vanilla JS, Vue, Svelte, Angular, SwiftUI, UIKit, Compose, MAUI, WinUI, Avalonia. The OpenAI HTTP surface is the only contract you depend on.
LangChain, Vercel AI SDK, OpenAI's official SDKs (Python, JS, Swift, .NET), Microsoft.SemanticKernel, CrewAI, LlamaIndex — all work unchanged. Point them at dvai.baseUrl.
Declarative loader for text + image + audio models (Gemma 4, LLaVA, Idefics, Whisper) on Transformers.js. Runs in the Web Worker by default; main thread stays unblocked during inference.
backend "auto" picks the best path per runtime — WebLLM in browsers, llama.cpp on mobile + desktop, Foundation Models on iOS 18.4+, etc. Override per request when you need to.
Zero API costs, zero server maintenance, no data leaves the device. Works offline. Suitable for healthcare, legal, finance, M&A, and any other regulated workload where cloud inference is a non-starter.
