{
  "version": 2,
  "weight_format": "affine",
  "profile": "JANG_8M",
  "source_model": {
    "name": "MiniCPM5-2B",
    "architecture": "llama (minicpm5)",
    "repo": "openbmb/MiniCPM5-2B",
    "revision": "3497c460c89e00520c3cfa2e73f49ab7647f1177",
    "note": "exact source revision this bundle was quantized from; upstream has since moved"
  },
  "has_vision": false,
  "has_audio": false,
  "tool_calling": {
    "dialect": "minicpm5_xml_function",
    "parser_module": "jang_tools.minicpm5.tool_parser",
    "tool_call_start": "<function name=\"",
    "tool_call_end": "</function>",
    "cdata_wrapped_values": true,
    "note": "mlx_lm's _infer_tool_parser matches NONE of the built-in literals against this template; without this parser tool calls are emitted as unparsed assistant text with no error."
  },
  "chat": {
    "sampling_defaults": {
      "temperature": 1.0,
      "top_p": 0.95,
      "do_sample": true,
      "source": "openbmb/MiniCPM5-2B model card + generation_config.json"
    },
    "eos_token_ids": [
      1,
      130073
    ],
    "thinking": {
      "supported": true,
      "template_flag": "enable_thinking",
      "default": "model decides (template appends nothing unless the flag is set)"
    }
  },
  "quantization": {
    "method": "affine",
    "quantization_backend": "mx.quantize",
    "modes": [
      "affine"
    ],
    "scale_dtype": "bfloat16",
    "norm_convention": "llama_rmsnorm_no_plus_one",
    "awq": {
      "applied": true,
      "alpha": 0.25,
      "statistic": "max|x| per input channel, geomean-normalized, clipped [0.5, 2.0]",
      "fold_sites": [
        "input_layernorm -> q/k/v",
        "post_attention_layernorm -> gate/up"
      ],
      "not_folded": [
        "o_proj (kv-head-tiled input)",
        "down_proj (nonlinear gate side)",
        "embed_tokens (lookup)"
      ],
      "norms_moved": 84
    },
    "gptq": {
      "applied": true,
      "objective": "||(deq(codes) - W) X^T||^2 on captured E[x x^T]",
      "guard": "best-of-RTN per tensor",
      "tensors_using_gptq": 295,
      "tensors_total": 296
    },
    "imatrix": {
      "applied": true,
      "source": "E[x_c^2] from the same capture"
    },
    "per_module_override_count": 296,
    "quantized_tensor_count": 296,
    "passthrough_tensor_count": 85
  },
  "runtime": {
    "total_weight_bytes": 2674327776,
    "total_weight_gb": 2.49,
    "bits_per_weight": 8.501
  },
  "capabilities": {
    "reasoning_parser": null,
    "tool_parser": "llama",
    "think_in_template": false,
    "supports_tools": true,
    "supports_thinking": false,
    "family": "llama",
    "modality": "text",
    "modalities": {
      "text": true,
      "vision": false,
      "audio": false,
      "video": false
    },
    "has_vision": false,
    "has_audio": false,
    "has_video": false,
    "cache_type": "kv"
  }
}