Skip to content
AI Model Radar

Nemotron Labs Audex 30B A3B

nvidia/Nemotron-Labs-Audex-30B-A3B
Open Source · chat · open-weights
GA Alert me on changes
Context
1M
Max output
—
Pricing
Open weights
Modalities
text
Released
06 Jul 2026
License: other · nvidia/Nemotron-Labs-Audex-30B-A3B
AI summary
● machine-written

NVIDIA Nemotron Labs Audex 30B A3B: unified audio-text LLM

Nemotron Labs Audex 30B A3B is a 30B MoE model with 3B activated parameters that extends text-only capabilities with audio understanding, speech recognition, text-to-speech, and audio generation. Built on Nemotron-Cascade-2, it supports up to 1M token context length and operates in both thinking and instruct modes. The model preserves reasoning and agentic capabilities while adding multimodal audio processing with marginal performance regression.

What's new
  • Unified audio-text architecture with discrete audio token vocabulary
  • Supports 1M-token context length
  • Both thinking and instruct (non-thinking) modes via <think> tags
  • Audio tasks: speech recognition, translation, text-to-speech, audio generation, speech-to-speech
  • Built on MoE backbone with 3B activated parameters from 30B total
Best for
Audio understanding and speech processingText reasoning and long-context agentic systemsSpeech-to-speech and audio generation tasksMultimodal enterprise agent systems
Sources

Source: https://huggingface.co/nvidia/Nemotron-Labs-Audex-30B-A3B