NVIDIA has unveiled **Audex (Nemotron-Labs-Audex-30B-A3B)**, a unified audio-text large language model that can both understand and generate audio and speech while retaining the text intelligence of its backbone. The model is released under a noncommercial license, along with a smaller Audex-2B variant.
Unlike many multimodal models that suffer a “text tax” (performance drops when adding audio or vision), Audex is designed to avoid such regression. It uses a simple single‑MoE Transformer decoder with 30B total parameters (3B activated per token) based on the Nemotron‑Cascade‑2‑30B‑A3B text‑only backbone. Audio inputs are projected into the text embedding space, and audio outputs are treated as text tokens, allowing standard training and inference stacks such as Megatron‑LM and vLLM to be used.
The system supports both an instruct mode and a thinking mode with a context length of up to 1M tokens. It employs a multi‑stage supervised fine‑tuning curriculum (text SFT, audio warmup, audio generation, audio understanding) combined with text‑only Cascade RL and multi‑domain on‑policy distillation to maintain text quality while adding audio capabilities. The training data mix includes 157.4B audio tokens and 320.5B text tokens across ASR, audio‑to‑text, TTS, text‑to‑audio, and audio understanding tasks.
In benchmarks, Audex matches or slightly exceeds its text‑only backbone on MMLU‑Redux, IMO AnswerBench, and several alignment and instruction‑following evaluations. It shows strong performance on speech recognition (6.82% WER on OpenASR) and leads among open models on MMAU, while also generating general audio beyond speech—a capability many open models lack. Practical inference recipes are provided for tasks such as audio understanding, ASR, TTS, and text‑to‑audio, with specific guidance on sampling, top‑k, and classifier‑free guidance.
The model is positioned as one of the strongest open alternatives for unified audio‑text generation, with strengths in minimal text regression, compatibility with standard LLM infrastructure, and general audio generation, though it remains non‑commercial and shows some gaps in certain audio‑understanding benchmarks.
*Source: NVIDIA has released Audex (Nemotron-Labs-Audex-30B-A3B), a unified audio-text LLM — NVIDIA Technical Blog, June 8, 2026.*
