argmaxinc

/

qwenasrkit-pro

Like 0

Follow

Argmax 224

Adds qwen3-asr-all_devices/, the Qwen3-ASR 1.7B tree packaged to run on the
A14-generation Neural Engine (M1 Macs) as well as on newer chips. Argmax SDK
serves it as Qwen3ASRVariant.qwen3ASR_1_7B_AllDevices (argmax-sdk-swift#480);
the default variant keeps downloading qwen3-asr/, which this PR does not touch.

Same layout as qwen3-asr/ (audio_encoder/1.7b + text_decoder/1.7b, 44 files,
1.79 GB), copied file for file from argmaxinc/qwenasrkit-pro-internal/qwen3-asr-kvcombined
(Hub commit f12b353), where it was staged and verified against its sources:

• audio_encoder/1.7b/*, Speculator.mlmodelc, the tokenizer files,
  decoder_manifest.json, LICENSE_NOTICE.txt: identical to the shipped qwen3-asr/ files.
• TextDecoderC0.mlmodelc / TextDecoderC1.mlmodelc: argmaxinc/axon-coreml
qwen3_asr/text_decoder_c{0,1}/1.7b-hf/W6A16-l1024-kvcombined-multifunction —
  the same weights with one combined KV state per layer (14 per chunk instead of 28),
  which is what fits the A14 generation's 26-state budget. Measured to load and run
  on an M1 Mac mini and an iPhone 12 mini; no measured cost on newer chips.
• TextDecoderEmbedHead.mlmodelc: argmaxinc/axon-coreml
qwen3_asr/text_decoder_embed_head/1.7b-hf/W6A16-multifunction. It has to be this
  head: C0 bakes a tether term derived from it and the shipped head's row differs
  (a mismatch decodes a lone EOS). In the SDK A/B, transcripts match the shipped
  folder on jfk and LibriSpeech and differ by one token on three of five clips.

Existing SDK releases are unaffected: they download qwen3-asr/ through the glob
*qwen3-asr/*, which does not match this folder.