---
license: cc-by-4.0
base_model: nvidia/canary-180m-flash
language:
- en
- de
- es
- fr
pipeline_tag: automatic-speech-recognition
tags:
- GGUF
- audio.cpp
---

# Canary 180M Flash GGUF

GGUF weights for [audio.cpp](https://github.com/0xShug0/audio.cpp), converted from
[NVIDIA Canary 180M Flash](https://huggingface.co/nvidia/canary-180m-flash).
The model recognizes English, German, Spanish, and French. Source and target languages
can be selected independently, matching the Python API. NVIDIA documents translation
quality for English-centric pairs.

| File | Weight format |
|---|---|
| `canary-180m-flash-f32.gguf` | Original tensor values and dtypes, predominantly F32 |
| `canary-180m-flash-q8_0.gguf` | Q8_0 for eligible tensors; other tensors retain supported precision |

Both files include the configuration and five tokenizers. Quantization may change
individual words; it does not guarantee identical output to the original weights.

```bash
audiocpp_cli --family canary_asr --task asr \
  --model canary-180m-flash-f32.gguf \
  --backend cuda --audio speech.wav --text-out transcript.txt --log
```

Long recordings are processed in chunks of up to 40 seconds. Use
`--request-option language=de --request-option target_language=en` for German to
English translation. This port provides offline text output, not timestamps.

The original model is by NVIDIA and licensed under
[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
The conversion changes the container; the Q8_0 variant additionally quantizes
eligible weights. Original checkpoints are not included here.