GLM-ASR-Nano-2512

Automatic Speech Recognition model from Z.ai (zai-org). Compact, open-source, beats Whisper V3 on multiple benchmarks.

Key facts

  • Type: Automatic Speech Recognition (ASR)
  • Architecture: Seq2Seq transformer
  • Params: 1.5B (BF16, ~2B size)
  • License: MIT
  • Languages: Mandarin Chinese, English, Cantonese (粤语), other dialects
  • Framework: Transformers (vLLM + SGLang support planned)

Features

  • Exceptional dialect support — optimized for Cantonese and regional variants
  • Robust for whisper/quiet (low-volume) speech
  • Outperforms OpenAI Whisper V3 on multiple benchmarks at smaller size

Benchmarks

  • Lowest average error rate (4.10) among comparable open-source models
  • Mean WER: 7.03
  • AMI WER: 16.15
  • RTFX: 145.28

Links: HuggingFace model · GitHub repo