{
  "_comment": "DeepSeek-Nano: Arquitetura DeepSeek-V3 escalonada para ~100k parâmetros ativos/token",

  "model_type": "deepseek_nano",
  "architectures": ["DeepSeekNanoForCausalLM"],
  "auto_map": {
    "AutoConfig":           "modeling_axion.AxionConfig",
    "AutoModelForCausalLM": "modeling_axion.DeepSeekNanoForCausalLM"
  },

  "vocab_size": 1024,

  "d_model": 64,
  "n_layers": 4,
  "n_heads": 4,
  "d_head": 16,

  "kv_lora_rank": 8,
  "q_lora_rank": 16,
  "rope_theta": 10000.0,
  "rope_scaling": null,

  "n_shared_experts": 1,
  "n_routed_experts": 4,
  "n_active_experts": 2,
  "d_ff": 64,
  "moe_aux_loss_coef": 0.0,
  "expert_bias_init": 0.0,

  "max_seq_len": 512,
  "dropout": 0.0,
  "norm_eps": 1e-6,
  "tie_embeddings": true,

  "pad_token_id": 0,
  "bos_token_id": 1,
  "eos_token_id": 2,
  "unk_token_id": 3,

  "_param_estimate": {
    "embedding": 65536,
    "per_layer_approx": 69632,
    "total_approx": 344064,
    "active_per_token": "~160k (MLA + 1 shared + 2 routed experts + norms)"
  }
}