NVIDIA has introduced Audex (Nemotron-Labs-Audex-30B-A3B), a unified audio-text large language model capable of understanding and generating both audio and speech. Unlike many multimodal models that experience a decline in text performance, Audex is designed to maintain the text intelligence of its Nemotron-Cascade-2-30B-A3B backbone. The model achieves this through a multi-stage training process and text-only reinforcement learning, resulting in competitive text benchmark scores and strong audio capabilities, including general audio generation beyond just speech. AI
IMPACT Sets a new standard for unified audio-text models by preserving text intelligence, potentially influencing future multimodal LLM development.
RANK_REASON Frontier-lab model release with system card [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- AF-Whisper
- Audex
- Audex-2B
- Audio Flamingo 3
- Mamba-Transformer
- Megatron-LM
- Nemotron-Cascade-2-30B-A3B
- Nemotron-Labs-Audex-30B-A3B
- NVIDIA
- Qwen3.5-35B-A3B
- Qwen3-Omni-30B-A3B-Thinking
- vLLM
- Whisper Large V3
- X-Codec
- X-Codec2
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
