Fun-Audio-Chat-8B

Fun-Audio-Chat-8B is a large audio language model published by Alibaba's FunAudioLLM team on 23 December 2025 under Apache 2.0, together with training and inference code. Unlike a transcription engine, it is meant to hold a conversation: audio goes in, and text or synthesised speech comes out, with function calling driven by spoken commands. Its central idea is what the authors call Dual-Resolution Speech Representations. Speech tokens normally run at 25 Hz while text tokens arrive at roughly 3 Hz, so a joint model spends most of its capacity on the faster stream and tends to forget what its text backbone knew. Here the shared backbone reads audio at 5 Hz, with tokens grouped five to one, while a separate refined head emits speech at the full 25 Hz. The producer reports that this halves the GPU hours needed for training. A second technique, Core-Cocktail training, is a two-stage fine-tune with an intermediate merge, aimed at the same forgetting problem. The name understates the download. The published checkpoint holds 9.45 billion parameters in bfloat16, not eight billion. The configuration file shows why: the text backbone matches Qwen3-8B exactly (36 layers, 4096 hidden size, 32 attention heads over 8 key-value heads, 151936-token vocabulary) and accounts for 8.19 billion of them. On top sits a 32-layer, 1280-wide audio encoder reading 128 mel bands in 30-second windows, and beside it a speech head with the geometry of Qwen3-0.6B. The figure in the name describes the part that thinks, not the parts that hear and speak — which is also why the producer asks for about 24 GB of GPU memory rather than the 16 GB an eight-billion model would suggest. One more practical point: speech output is not self-contained. The quick-start instructions tell you to download a second checkpoint, Fun-CosyVoice3-0.5B-2512, alongside the main model. The technical report describes three members of the family — this 8B model, a mixture-of-experts 30B-A3B and a full-duplex variant able to listen and speak at once. Only the 8B is open. Languages are limited to English and Chinese, so a Polish-speaking user has no path here. The producer reports leading results among models of comparable size on spoken question answering (OpenAudioBench, VoiceBench, UltraEval-Audio), audio understanding (MMAU, MMAU-Pro, MMSU), speech function calling (Speech-ACEBench, Speech-BFCL, Speech-SmartInteract) and spoken instruction following (VStyle).

▸News

In voice models, the number in the name counts only the part that thinks

9/5/2026 · Hugging Face — metadane plików wag (safetensors) i pliki config.json modeli głosowych

▸Videos

No videos yet.

▸Reviews