Argmax gets Qwen3-ASR running inside the iPhone, with speech transcribed on the device
Pro SDK 3 has reached a stable release. The multilingual recognition model runs on the phone rather than a server, and the company has opened a TestFlight app for trying it without writing code.
Argmax, which builds AI infrastructure for Apple hardware, has shipped the stable release of Pro SDK 3 (3.0.6). The central claim is that multilingual speech recognition now runs on the iPhone's own silicon.
The model is Qwen3-ASR-1.7b, the 1.7-billion-parameter version of Alibaba's open speech recognition model, which Argmax has ported to Apple Silicon and packaged into the SDK. Audio does not leave for a server; the work finishes on the phone. That matters for privacy, and equally for working without a connection: on a plane, in a tunnel, out of coverage, it behaves the same.
The SDK's real-time API provides custom vocabulary and word-level timestamps. Custom vocabulary can be changed at runtime, which means a meeting app can teach it participants' names, or a clinical app drug names, while someone is still speaking. The company gives a lips-to-screen latency figure of 160 milliseconds; that measurement is Argmax's own.
Trying it requires no code: the Argmax Playground app opens on iPhone and Mac through TestFlight.
Details are on the Argmax blog, with the open-source side on GitHub.
