Argmax SDK 2

Just in time for the holidays, we are announcing Argmax SDK 2 in early access!

‍

‍
Real-time Transcription gets Speakers

‍We brought frontier-level speaker diarization accuracy with real-time mode into Argmax SDK 2 powered by Nvidia Sortformer. We have been using this system to transcribe our internal meetings at Argmax and we have observed dramatically improved speaker consistency and reduced speaker count errors compared to the previous generation of Argmax SDK.

In our benchmark preview, Argmax SDK 2's Real-time Transcription with Speakers surpasses top cloud APIs like Deepgram in speaker-attributed transcription accuracy as measured by cpWER (lower the better) on the callhome (telephone conversations) dataset. As usual, we will publish reproducible and open-source final benchmarks in OpenBench along with the general availability of Argmax SDK 2 in February 2026. Here is a preview:

‍

‍

Open-sourcing Argmax SpeakerKit

Argmax SDK 2, powered by Nvidia Sortformer, brings about a generational accuracy leap for speaker diarization compared to Argmax SDK 1's SpeakerKit, powered by pyannote 4. Hence, we will be open-sourcing SpeakerKit's pyannote 4 engine alongside the general availability of Argmax SDK 2.

Argmax is committed to democratizing on-device AI with free and open-source developer tools while building frontier-level capabilities into our commercial SDK for developers and Enterprises with the most demanding requirements.

‍

Early Access

If you are eager to apply for early access ahead of the planned February 2026 general availability, please send us a note with your use case at early-access@argmaxinc.com!