Audio Flamingo Next: Next-Generation Open Audio-Language Models for Speech, Sound, and Music

A consortium of researchers from NVIDIA, the University of Maryland, and KAIST has unveiled Audio Flamingo Next (AF-Next), positioning it as the most capable model in its open-source series for understanding speech, sound, and music. The model represents a significant leap from its predecessor, Audio Flamingo 3, by introducing a stronger foundational architecture and, critically, the ability to handle long and complex audio inputs of up to 30 minutes. To achieve this, the team conducted a systematic analysis of previous limitations and then curated and scaled new datasets totaling over 1 million hours, expanding existing resources like AudioSkills-XL and LongAudio-XL.

A key technical innovation is the Temporal Audio Chain-of-Thought, a new reasoning paradigm that explicitly grounds intermediate reasoning steps to specific timestamps within long audio. This allows for fine-grained temporal alignment, improving both the model's accuracy on complex tasks and the interpretability of its outputs. The model was trained using a multi-stage curriculum spanning pre-training, mid-training, and post-training phases.

Extensive evaluation across 20 audio understanding and reasoning benchmarks shows AF-Next outperforming similarly sized open models by large margins. The paper notes it remains highly competitive with, and sometimes surpasses, much larger open-weight and closed models. Beyond benchmark performance, AF-Next demonstrates strong real-world utility and generalization to unseen tasks. The researchers are open-sourcing three variants of the model: AF-Next-Instruct, AF-Next-Think, and AF-Next-Captioner, along with all data, code, and methods.