Higgs Audio is an expressive audio foundation model from Boson AI, pretrained on over 10 million hours of audio and text data, capable of generating natural multi-speaker dialogues, melodic humming with cloned voices, and simultaneous speech with background music. The latest V2.5 release condenses the architecture to 1B parameters while surpassing the prior 3B model in speed and accuracy through Group Relative Policy Optimization (GRPO) alignment on a curated Voice Bank dataset. It achieves state-of-the-art results on EmergentTTS-Eval, outperforming GPT-4o-mini-TTS on expressive emotion and intonation tasks.