{
  "architectures": [
    "NexaVisionMoEForConditionalGeneration"
  ],
  "auto_map": {
    "AutoConfig": "configuration_nexa.NexaVisionMoEConfig",
    "AutoModelForCausalLM": "modeling_nexa.NexaVisionMoEForConditionalGeneration"
  },
  "model_type": "nexa_vision_moe",
  "transformers_version": "5.12.1",
  "vision_model_id": "google/siglip2-so400m-patch16-naflex",
  "llm_model_id": "Neura-Tech-AI/Nexa-AI-4x4B-Instruct",
  "vision_dim": 1152,
  "llm_dim": 2560,
  "hidden_size": 2560,
  "intermediate_size": 9728,
  "moe_intermediate_size": 9728,
  "num_local_experts": 4,
  "num_experts": 4,
  "num_experts_per_tok": 2,
  "num_hidden_layers": 36,
  "num_attention_heads": 32,
  "num_key_value_heads": 8,
  "head_dim": 128,
  "hidden_act": "silu",
  "max_position_embeddings": 262144,
  "vocab_size": 151936,
  "bos_token_id": 151643,
  "eos_token_id": 151645,
  "pad_token_id": null,
  "tie_word_embeddings": true,
  "rms_norm_eps": 1e-06,
  "initializer_range": 0.02,
  "attention_bias": false,
  "attention_dropout": 0.0,
  "decoder_sparse_step": 1,
  "dtype": "bfloat16",
  "norm_topk_prob": true,
  "output_router_logits": false,
  "router_aux_loss_coef": 0.001,
  "use_cache": true,
  "use_sliding_window": false,
  "sliding_window": null,
  "max_window_layers": 36,
  "mlp_only_layers": [],
  "rope_parameters": {
    "rope_theta": 5000000,
    "rope_type": "default"
  },
  "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",
    "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",
    "full_attention",
    "full_attention"
  ]
}