TangoFlux

Text prompts become 44.1 kHz audio in roughly 3.7 seconds with TangoFlux, a 515M parameter text-to-audio model from the DeCLaRe Lab at Singapore University of Technology and Design, produced with NVIDIA and Lambda and appearing at ICLR 2026. The architecture stacks FluxTransformer blocks, Diffusion Transformer and Multimodal Diffusion Transformer layers conditioned on text and duration embeddings, trained with flow matching over rectified flow trajectories; a final CRPO stage, CLAP-Ranked Preference Optimization, builds preference pairs ranked by the CLAP model. Generating up to 30 seconds of audio takes about 3.7 seconds on a single NVIDIA A40, with 25 sampling steps the default and 50 recommended for higher quality. The repository installs via pip from GitHub and provides a CLI, Python API, Gradio demo, and a ComfyUI integration, with weights hosted on Hugging Face. Licensing is the main caveat: the model falls under the Stability AI Community License with WavCaps-derived data restricted to academic use, making it research-only, so commercial deployment needs separate licensing. A GPU is effectively required.