Fun-Audio-Chat-8B

Fun-Audio-Chat-8B is a large audio language model from FunAudioLLM (flink-jp) released in December 2025 with Apache 2.0 licensing, training code, and inference code. Its core innovation is dual-resolution speech representations: audio is compressed to a 5Hz frame rate for LLM processing, cutting GPU hours by nearly half, while a parallel speech head regenerates high-quality 25Hz speech tokens.

The model supports spoken question answering, audio understanding, speech function calling, speech instruction following, and voice empathy, with built-in emotion detection. It comes in dense 8B and 30B-A3B MoE scales, and is presented in a technical report with extensive VoiceBench, UltraEval-Audio, and internal benchmark comparisons.

Language coverage is currently English and Chinese, and local inference needs around 24GB of VRAM. It is a strong open candidate for self-hosted voice assistants where data cannot leave the premises.