f32/metadata.json

{
  "model_id": "Qwen/Qwen3-ASR-0.6B",
  "architecture": "Qwen3ASRForConditionalGeneration",
  "sample_rate": 16000,
  "num_mel_bins": 128,
  "max_audio_seconds": 30.0,
  "max_seq_length": 4096,
  "audio_encoder": {
    "n_window": 50,
    "n_window_infer": 800,
    "mel_window_size": 100,
    "conv_downsample_factor": 8,
    "d_model": 896,
    "output_dim": 1024,
    "num_layers": 18,
    "num_heads": 14
  },
  "text_decoder": {
    "hidden_size": 1024,
    "intermediate_size": 3072,
    "num_layers": 28,
    "num_attention_heads": 16,
    "num_kv_heads": 8,
    "head_dim": 128,
    "vocab_size": 151936,
    "rope_theta": 1000000,
    "mrope_section": [24, 20, 20]
  },
  "special_tokens": {
    "audio_start_token_id": 151669,
    "audio_end_token_id": 151670,
    "audio_token_id": 151676,
    "im_start_token_id": 151644,
    "im_end_token_id": 151645,
    "system_token_id": 8948,
    "user_token_id": 872,
    "assistant_token_id": 77091,
    "newline_token_id": 198,
    "eos_token_ids": [151645, 151643]
  },
  "components": {
    "audio_encoder": {
      "path": "qwen3_asr_audio_encoder.mlpackage",
      "precision": "float16"
    },
    "embedding": {
      "path": "qwen3_asr_embedding.mlpackage",
      "precision": "float16"
    },
    "decoder_stack": {
      "path": "qwen3_asr_decoder_stack.mlpackage",
      "num_layers": 28,
      "weight_precision": "float32",
      "compute_precision": "float32",
      "note": "All 28 decoder layers consolidated into one model with stacked KV caches. Full float32 weights and compute."
    },
    "lm_head": {
      "path": "qwen3_asr_lm_head.mlpackage",
      "weight_precision": "float32",
      "compute_precision": "float32"
    }
  },
  "export_settings": {
    "compute_units": "CPU_ONLY",
    "deployment_target": "iOS17"
  }
}