Canary-Qwen-2.5B

NVIDIA's open-weights English speech recognition model that pairs a FastConformer encoder with a Qwen3-1.7B LLM decoder.

Where it ranks

B
Transcription
Models · Local · 4.3% on Artificial Analysis Speech to Text: AA-WER v2 (non-streaming)

Vendor: NVIDIA
Family: Canary
Released: 17 Jul 2025
Weights: Open: download and run it yourself
License: CC-BY-4.0
Parameters: 2.56B
GPU memory to run locally: ~6 GB (Derived: 2.56B parameters (Hugging Face safetensors count) x 2 bytes = ~5.1 GB of bf16 weights.)
API price: $0.00074 per minute of audio (Artificial Analysis lists $0.74 per 1,000 minutes (Canary Qwen 2.5B, NVIDIA, provider Replicate); divided by 1,000.)
Access: Free (open weights)
Last verified: 18 Oct 2026

Price from artificialanalysis.ai, 18 Oct 2026.

Benchmark	Result	Entry on the leaderboard	As of	Source
Artificial Analysis Speech to Text: AA-WER v2 (non-streaming)	4.3%	`Canary Qwen 2.5B, NVIDIA` AA-WER v2 (non-streaming), provider Replicate; embedded value 0.0427 converted to percent and rounded; page shows no last-updated date.	18 Oct 2026	artificialanalysis.ai
Hugging Face Open ASR Leaderboard (English, short-form average WER)	5.9%	`nvidia/canary-qwen-2.5b` Average WER over the board's default datasets (English short-form, version 17-10-2026); rank 18; read from the Space's initial table data.	17 Oct 2026	huggingface.co