Fun-ASR-Nano 2512

Fun-ASR-Nano is the open-weight end of Alibaba's December 2025 transcription release from Tongyi Lab. It ships as a pair of 800-million-parameter models published together under Apache 2.0, and the split between them matters more than the shared name suggests. The plain Nano is the deep one: Chinese, English and Japanese only, but with seven major Chinese dialects — Wu, Cantonese, Min, Hakka, Gan, Xiang, Jin — and twenty-six regional accents, trained on what the producer describes as tens of millions of hours of real speech. The MLT variant is the broad one: thirty-one languages including Polish, trained on hundreds of thousands of hours, roughly two orders of magnitude less material per language. The model is tuned for conditions that defeat general transcribers. On the producer's industry test sets it reports 5.79 per cent word error in far-field audio against 22.21 for Whisper-large-v3, 28.18 on dialect speech against 66.14, and 30.85 on song lyrics against 54.82 — a category most transcribers decline outright. On the standard public benchmarks the gap closes to a draw or slightly worse: 1.76 on LibriSpeech clean against Whisper's 1.86, 4.33 on LibriSpeech other against 3.43. It runs streaming with low latency, and since mid-2026 also as a self-contained GGUF binary on a processor without Python or a graphics card. Two honest notes. The published benchmark tables place the small open model beside a 7.7-billion-parameter Fun-ASR that wins nearly every row — that larger model is marked as not open source and is reachable only as a hosted service, so the figures a reader is most likely to quote are not the figures they can download. And the model's own labels on Hugging Face advertise timestamps and speaker diarisation, while the to-do list further down the same card records both as not yet supported.

▸Videos

No videos yet.

▸Reviews