NVIDIA releases Audex 30B for audio-text AI agents
NVIDIA’s Nemotron-Labs-Audex-30B-A3B is a that handles both audio and text. It is built on Nemotron-Cascade-2-30B-A3B, a text-only base model with 30 billion total parameters and 3 billion used at a time. The model adds an for speech and general audio input, plus audio tokens for speech and other sound output.
It can work on audio understanding, , speech translation, , audio generation, and speech-to-speech generation. It is designed to keep the base model’s reasoning, knowledge, alignment, long-context handling, and agent abilities with little or no drop. It supports both a and a direct instruct mode.
It follows the ChatML format and can handle up to a 1 million-token context.
Key points
- Audex-30B-A3B combines audio and text in one .
- It has 30 billion total parameters, but uses 3 billion at a time through a MoE design.
- It supports audio understanding, , translation, , audio generation, and speech-to-speech generation.
- It aims to keep strong reasoning, long-context handling, knowledge, and agent abilities from its text-only base model.
- It supports up to a 1 million-token context, which could help agents work with long inputs.