Run AI on Your Users’ Devices

Most on-device features start with an off-the-shelf model and then try to make it fit, while we start with the task and go from there. We start with either distilling a large model or fine tuning a small one and then quantizing it to fit memory and performance constraints. Where that isn't enough, we also prune before quantizing. Alongside model development we create a pipeline with benchmarks designed for measuring memory usage, latency, thermal throttling and battery draw, so you know how it will perform in the real world. You get the model weights, export pipeline and benchmarks to keep and build on.

We help teams bring AI directly into their mobile apps. Our open-source React Native ExecuTorch library provides typed React Native API for LLMs, speech-to-text, text-to-speech, OCR, embeddings and object detection. We can ship cross platform with XNNPACK, or use platform-specific acceleration: MLX or CoreML for iOS and Vulkan for Android. We have experience with deploying models like Whisper, Kokoro, LFM2.5, Gemma and Yolo on real phones. Saying that, we can bring the same work to native iOS and Android applications if React Native isn’t part of your stack.

On-device AI isn’t limited to phones, and we help teams run the same kinds of models on laptops and in browsers. We build model runtimes using TypeGPU, our type-safe WebGPU library, so inference can happen directly on the user’s machine without another app to install or a server to send data to. We take care of adapting and optimizing the model for the hardware available in the browser or desktop environment, rather than assuming every machine can run the same thing.

A model on its own rarely makes a useful feature, so we can build the rest of the pipeline around it for you. With React Native RAG, for example, retrieval-augmented generation can run entirely locally, including embeddings, vector search and generation, while our agent pipelines can call a cloud model only when there’s a good reason to. When keeping data local is a regulatory requirement rather than a performance choice, we can also help you with privacy filters that will identify and remove PII before it reaches any model.

Sometimes the job is not just getting a model to run, but turning it into a product that people can actually use. We’ve been building production mobile and web applications since 2012, and today we combine that product experience with our work in on-device AI. We can handle the interface, model downloads and storage, device compatibility, fallbacks and release process alongside the AI itself. Private Mind is a great example – it’s our AI assistant that runs entirely on the phone.

If you already have something running but it’s too slow, too large, or you’re no longer sure that running locally is the right choice, we can take a look at the whole setup. We review the model, runtime, quantization and latency, then benchmark it on the devices that actually matter for your users, so the recommendations are based on real performance rather than theoretical specifications. You get a prioritized set of changes and a clear assessment of whether on-device still makes sense, and we can either implement the improvements ourselves or work alongside your engineers to get them done.