config.json

{
  "model_type": "needle",
  "architectures": ["NeedleForToolCalling"],
  "name_or_path": "Cactus-Compute/needle2",
  "vocab_size": 8192,
  "hidden_size": 512,
  "num_hidden_layers": 27,
  "num_attention_heads": 8,
  "num_key_value_heads": 4,
  "head_dim": 64,
  "max_position_embeddings": 2048,
  "rope_theta": 100000.0,
  "tie_word_embeddings": true,
  "torch_dtype": "float16",
  "total_parameters": 44900000,
  "sliding_window": 256,
  "attention": "gqa",
  "mlp": "hadamard",
  "extras": {
    "engram": {"sites": [2, 15], "orders": [2, 3], "slots": 8192, "sub_dim": 128, "conv_taps": 4, "conv_dilation": 3},
    "mhc_lanes": 4,
    "heads": ["contrastive_retrieval", "confidence"]
  },
  "quantization": {
    "method": "cactus-quants",
    "scheme": "embedding=4,mhc=4,default=2",
    "effective_bits": 2.2,
    "group_size": 128,
    "kv_cache_bits": 8,
    "activation_bits": 8
  },
  "deployment": {
    "format": "cact",
    "binary_size_mb": 14,
    "session_ram_mb": 28,
    "runtime": "self-contained C++ / WebAssembly, no dependencies"
  }
}