{
  "architectures" : [
    "Qwen3ForCausalLM"
  ],
  "attention_bias" : false,
  "attention_dropout" : 0,
  "bos_token_id" : 3,
  "dtype" : "bfloat16",
  "eos_token_id" : 3,
  "head_dim" : 128,
  "hidden_act" : "silu",
  "hidden_size" : 512,
  "initializer_range" : 0.041666666666666664,
  "intermediate_size" : 2304,
  "layer_types" : [
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention",
    "full_attention"
  ],
  "max_position_embeddings" : 1024,
  "max_window_layers" : 17,
  "model_type" : "soprano",
  "num_attention_heads" : 4,
  "num_hidden_layers" : 17,
  "num_key_value_heads" : 1,
  "quantization" : {
    "bits" : 4,
    "group_size" : 64,
    "mode" : "affine"
  },
  "quantization_config" : {
    "bits" : 4,
    "group_size" : 64,
    "mode" : "affine"
  },
  "rms_norm_eps" : 9.9999999999999995e-07,
  "rope_scaling" : null,
  "rope_theta" : 10000,
  "sliding_window" : null,
  "tie_word_embeddings" : false,
  "transformers_version" : "4.57.0",
  "use_cache" : true,
  "use_sliding_window" : false,
  "vocab_size" : 8192
}