{
  "model_file": "Meltdown_Q8.gguf",
  "description": "Meltdown APE agent model — Q8_0 quantized GGUF for local tool-calling harnesses",
  "base_model": "Qwen2.5-Coder-3B-Instruct (fine-tuned, exported Q8_0)",
  "quantization": "q8_0",
  "multimodal": false,
  "inference": {
    "ctx_size": 32000,
    "max_output_tokens": 16384,
    "cache_type_k": "q8_0",
    "cache_type_v": "q8_0",
    "gpu_layers": -1,
    "threads": 4,
    "temperature": 0.2,
    "top_p": 0.95,
    "top_k": 40
  },
  "harness": {
    "system_prompt_file": "system_prompt.txt",
    "append_tool_catalog_at_runtime": true,
    "append_sandbox_context_at_runtime": true,
    "tool_result_header_format": "[Tool result — {tool_name}]:",
    "max_agent_rounds": 24,
    "stop_after_tool_call_close_tag": true
  },
  "notes": [
    "Use temperature 0.1–0.3 for reliable APE tags and tool_call JSON.",
    "Append tool list and sandbox paths to the system prompt at session start.",
    "max_output_tokens should be >= 16384 during agent loops to avoid truncated tool calls.",
    "gpu_layers: -1 means offload all layers (llama.cpp convention); adjust to your VRAM."
  ]
}