Downloads · 30 days
20
18% of all-time downloads
qwen3-asr-0.6b-onnx
MBFSAITeam/qwen3-asr-0.6b-onnx
qwen3-asr-0.6b-onnx is a automatic speech recognition model from MBFSAITeam. Use it when you need speech turned into text. The card lists the license as apache-2.0.
ONNX export of Qwen/Qwen3-ASR-0.6B optimized for the MBFS AI Team
All-time downloads
112
Public
Repo size
6.8 GB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.data5.4 GB · 80%
At a glance
- Task
- Automatic Speech Recognition
- License
- apache-2.0
- Access
- Public
- Created
- Jun 8, 2026
- Updated
- Jun 8, 2026
- SHA
- 16a13c0c
Base models
Model card
From the Hugging Face model README
Qwen3-ASR-0.6B — ONNX Export (mbfs-stt)
ONNX export of Qwen/Qwen3-ASR-0.6B optimized for the MBFS AI Team
Bundle contents
| File | Shape / Description |
|---|---|
encoder.onnx | mel [1,128,3000] → audio_features [1,A,1024] — fixed 30 s window |
decoder_init.onnx | input_ids, position_ids, audio_features → logits, present_keys, present_values — dynamic seq |
decoder_step.onnx | input_embeds, position_ids, past_keys, past_values → logits, present_keys, present_values — dynamic KV cache |
embed_tokens.bin | Raw row-major [V, H] weight (float32 or float16) |
config.json | hidden_size, vocab_size, embed_tokens_shape, embed_tokens_dtype |
tokenizer.json | HuggingFace fast tokenizer |
Architecture notes
- Encoder: TorchScript ONNX export, fixed at 3000 mel frames (30 s) due to data-dependent control flow in the audio tower. Input must be zero-padded to exactly 3000 frames.
- Decoder (init + step): Dynamo (
torch.export) export with truly dynamic sequence and KV-cache dimensions. KV cache uses stacked layout[L, B, KVH, S, D](28 decoder layers). - Audio scatter offset: fixed constant
9(prompt prefix length). - Opset: 17
Usage (Rust engine)
# config.yml
asr:
engine_type: qwen_asr
qwen_model_path: models/qwen3-asr-0.6b
Limitations
- Encoder is hard-coded to 30 s chunks. Shorter audio must be zero-padded to 480 000 samples before mel computation.
- Float32 weights by default; use --emb-dtype float16 for a smaller embed_tokens.bin.
- Task
- Automatic Speech Recognition
- License
- apache-2.0
- Languages
- multilingual
- Created
- Jun 8, 2026
- Updated
- Jun 8, 2026