Qwen3-ASR#

Qwen3-ASR is an audio speech recognition model from the Qwen3 family. Megatron Bridge supports checkpoint conversion and audio inference through the Qwen3-ASR bridge.

Supported Variants#

  • Qwen3-ASR-1.7B: https://huggingface.co/Qwen/Qwen3-ASR-1.7B

Architecture Notes#

  • Uses the Qwen3ASRForConditionalGeneration architecture on the Hugging Face side.

  • The Megatron implementation includes a Qwen3-ASR thinker model, transformer config, and ASR-specific RoPE support.

  • Example conversion and round-trip validation use --trust-remote-code.

Examples#

For checkpoint import/export, round-trip validation, and audio inference commands, see the Qwen3-ASR examples README.