Qwen3-ASR-1.7B

Model

Qwen3-ASR-1.7B — Qwen3VL-based Audio Speech Recognition model for transcribing English speech audio.

• Audio encoder: 24-layer Transformer encoder (hidden=1024, 16 heads, head_dim=64, FFN=4096, output_dim=2048) with 3 initial Conv2D layers (3×3, 480 channels)
• Text decoder: 28-layer Qwen3VL decoder (hidden=2048, 16Q/8KV heads via GQA, head_dim=128, FFN=6144, SwiGLU, QK-Norm, M-RoPE, vocab 151 936)
• Total parameters: ~1.7B

Reference implementation: reference-llm-models / qwen3-asr

---

Available weights

The model consists of two weight files — decoder (LLM backbone) and encoder (audio encoder).

Component	Directory	File	Dtype	Tensors
Decoder	Qwen3_ASR_1.7B_fp16_artifacts/	qwen3_asr_decoder_fp16.npz	FP16	311
Decoder	Qwen3_ASR_1.7B_fp16_artifacts/	Qwen3-ASR-1.7B-FP16.gguf	FP16	311
Decoder	Qwen3_ASR_1.7B_fp32_artifacts/	qwen3_asr_decoder_fp32.npz	FP32	311
Decoder	Qwen3_ASR_1.7B_fp32_artifacts/	Qwen3-ASR-1.7B-FP32.gguf	FP32	311
Encoder	Qwen3_ASR_1.7B_fp16_artifacts/	qwen3_asr_encoder_fp16.npz	FP16	398
Encoder	Qwen3_ASR_1.7B_fp16_artifacts/	mmproj-Qwen3-ASR-1.7b-FP16.gguf	FP16	398
Encoder	Qwen3_ASR_1.7B_fp32_artifacts/	qwen3_asr_encoder_fp32.npz	FP32	398
Encoder	Qwen3_ASR_1.7B_fp32_artifacts/	mmproj-Qwen3-ASR-1.7b-FP32.gguf	FP32	398

Both NPZ files must be present for inference — the decoder NPZ contains the language model (311 tensors: embed_tokens, 28× transformer layers, final norm, lm_head), and the encoder NPZ contains the audio encoder (397 tensors: 3× conv2d, 24× transformer layers, projection head).

---

How to run

set USE_TORCH=1
python generate.py ^
    --audio sample.wav ^
    --decoder-weights ./Qwen3_ASR_1.7B_fp16_artifacts/qwen3_asr_decoder_fp16.npz ^
    --encoder-weights ./Qwen3_ASR_1.7B_fp16_artifacts/qwen3_asr_encoder_fp16.npz ^
    --hf-model-path ../Qwen3-ASR-1.7B ^
    --verbose

Audio feature validation only (no decoding)

set USE_TORCH=1
python generate.py ^
    --audio sample.wav ^
    --decoder-weights ./Qwen3_ASR_1.7B_fp16_artifacts/qwen3_asr_decoder_fp16.npz ^
    --encoder-weights ./Qwen3_ASR_1.7B_fp16_artifacts/qwen3_asr_encoder_fp16.npz ^
    --hf-model-path ../Qwen3-ASR-1.7B ^
    --validate-audio-features ^
    --verbose

Dump intermediates for numerical comparison

set USE_TORCH=1
python generate.py ^
    --audio sample.wav ^
    --decoder-weights ./Qwen3_ASR_1.7B_fp16_artifacts/qwen3_asr_decoder_fp16.npz ^
    --encoder-weights ./Qwen3_ASR_1.7B_fp16_artifacts/qwen3_asr_encoder_fp16.npz ^
    --hf-model-path ../Qwen3-ASR-1.7B ^
    --intermediate-output-path intermediates.npz ^
    --verbose

---

Key configuration

Config file: model_config.json

Parameter	Value
DECODER_NUM_HIDDEN_LAYERS	28
DECODER_HIDDEN_SIZE	2048
DECODER_INTERMEDIATE_SIZE (FFN)	6144
DECODER_NUM_ATTENTION_HEADS	16
DECODER_NUM_KEY_VALUE_HEADS	8
DECODER_HEAD_DIM	128
VOCAB_SIZE	151936
DECODER_RMS_NORM_EPS	1e-6
DECODER_ROPE_THETA	1000000.0
DECODER_MAX_SEQ_LEN	1024
DECODER_MROPE_SECTION	[24, 20, 20]
DECODER_ENABLE_ROPE	1
ENCODER_NUM_LAYERS	24
ENCODER_HIDDEN_SIZE	1024
ENCODER_NUM_HEADS	16
ENCODER_HEAD_DIM	64
ENCODER_FFN_DIM	4096
ENCODER_OUTPUT_DIM	2048
ENCODER_LAYER_NORM_EPS	1e-5
ENCODER_MAX_SOURCE_POSITIONS	1500
ENCODER_DOWNSAMPLE_HIDDEN_SIZE	480
ENCODER_NUM_MEL_BINS	128
ENCODER_N_WINDOW	50
ENCODER_N_WINDOW_INFER	800
ENCODER_CONV_CHUNKSIZE	500

Controlled via model_config.json (read by llama_model.py and llama_model_audio.py).

Note: set USE_TORCH=1 is mandatory. The reference code is based on torch_extend_ops.py which requires PyTorch with CUDA.

---

Input / Output

Input

• Audio: path to a local .wav file (--audio), 16 kHz mono
• Dtype: FP16 (default). The reference weights are BF16/FP16; an FP32 variant is also available for higher-precision comparison.

Output

Loading decoder weights from qwen3_asr_decoder.npz …
Loading encoder weights from qwen3_asr_encoder.npz …
Transcript: Oh yeah, yeah. But you know, it's not a big deal..

The output is a plain-text transcript of the spoken audio.