Open4bits/Nanbeige4.1-3B-mlx-4Bit

1{2    "architectures": [3        "LlamaForCausalLM"4    ],5    "attention_bias": false,6    "attention_dropout": 0.0,7    "bos_token_id": 166100,8    "dtype": "bfloat16",9    "embd_pdrop": 0.0,10    "eos_token_id": 166101,11    "head_dim": 128,12    "hidden_act": "silu",13    "hidden_size": 2560,14    "initializer_range": 0.02,15    "intermediate_size": 10496,16    "max_position_embeddings": 262144,17    "mlp_bias": false,18    "model_type": "llama",19    "num_attention_heads": 20,20    "num_hidden_layers": 32,21    "num_key_value_heads": 4,22    "pad_token_id": 0,23    "pretraining_tp": 1,24    "quantization": {25        "group_size": 64,26        "bits": 4,27        "mode": "affine"28    },29    "quantization_config": {30        "group_size": 64,31        "bits": 4,32        "mode": "affine"33    },34    "resid_pdrop": 0.0,35    "rms_norm_eps": 1e-05,36    "rope_scaling": null,37    "rope_theta": 70000000,38    "tie_word_embeddings": false,39    "torch_dtype": "bfloat16",40    "transformers_version": "4.51.3",41    "use_cache": true,42    "vocab_size": 16614443}