config.json

{
  "model_type": "needle",
  "engine_version": "3.1.0",
  "architectures": [
    "NeedleForToolCalling"
  ],
  "name_or_path": "Cactus-Compute/needle3",
  "vocab_size": 8192,
  "output_vocab_size": 8192,
  "hidden_size": 768,
  "num_hidden_layers": 20,
  "num_attention_heads": 12,
  "num_key_value_heads": 2,
  "qk_head_dim": 48,
  "value_head_dim": 64,
  "max_position_embeddings": 8192,
  "rope_theta": 100000.0,
  "tie_word_embeddings": true,
  "total_parameters": 121021910,
  "sliding_window": 1024,
  "attention": "gqa",
  "mlp": "hadamard",
  "extras": {
    "engram": {
      "sites": [
        3,
        7,
        11,
        15,
        19
      ],
      "orders": [
        2,
        3
      ],
      "slots": 18432,
      "sub_dim": 128
    },
    "global_attention_layers": [
      4,
      9,
      14,
      19
    ],
    "qkv_conv_taps": 3,
    "mhc_lanes": 4,
    "heads": [
      "confidence"
    ],
    "modalities": [
      "text"
    ]
  },
  "quantization": {
    "method": "cactus-quants",
    "scheme": "embedding=4,mhc=4,default=2",
    "group_size": 128,
    "kv_cache_bits": 8,
    "activation_bits": 8
  },
  "deployment": {
    "format": "cact",
    "runtime": "self-contained C++ / WebAssembly, no dependencies"
  },
  "kv_window": 256
}