Canary 180M Flash

Canary-180M-Flash is the smallest variant of NVIDIA's Canary family of encoder-decoder speech models: a 182M-parameter multitask model with a 17-layer FastConformer encoder and a 4-layer transformer decoder. It does offline ASR in English, German, Spanish, and French plus speech translation between English and the other three languages, in about 208 MB at Q8 quantization. It trades the broader 25-language coverage of Canary-1B-v2 for speed and footprint, making it suitable for on-device and CPU-friendly transcription.

Weights are published on Hugging Face under CC-BY-4.0 with NeMo tooling, and the model is widely supported in third-party runtimes (sherpa-onnx, transcribe.cpp, llama.cpp-adjacent ports). Exact release date is era-level (spring 2024, alongside the other Canary flash variants) and flagged as such.