{
  "schemaVersion": 1,
  "count": 13,
  "runtimes": [
    {
      "id": "llama-cpp",
      "displayName": "llama.cpp",
      "formats": [
        "gguf"
      ],
      "toolchain": "Bundled with the OpenEng runtime — no external install. CPU everywhere; GPU offload via Metal, CUDA, or Vulkan when present.",
      "blurb": "The GGUF-native inference engine. The default on-device path for quantized models."
    },
    {
      "id": "onnx-runtime",
      "displayName": "ONNX Runtime",
      "formats": [
        "onnx"
      ],
      "toolchain": "Bundled with the OpenEng runtime — no external install. CPU execution provider by default; hardware EPs when available.",
      "blurb": "Cross-platform inference for ONNX graphs — the path for bundled utility models."
    },
    {
      "id": "vllm",
      "displayName": "vLLM",
      "formats": [
        "safetensors",
        "gguf",
        "gptq",
        "awq"
      ],
      "toolchain": "Python 3.10+ with a CUDA-capable GPU; installed via pip. Runs as a local OpenAI-compatible server.",
      "blurb": "High-throughput PagedAttention serving engine for safetensors and quantized (GPTQ/AWQ) checkpoints."
    },
    {
      "id": "sglang",
      "displayName": "SGLang",
      "formats": [
        "safetensors"
      ],
      "toolchain": "Python with a CUDA-capable GPU; installed via pip. Runs as a local serving process.",
      "blurb": "RadixAttention serving engine with structured-generation support for safetensors checkpoints."
    },
    {
      "id": "ex-llama-v2",
      "displayName": "ExLlamaV2",
      "formats": [
        "gptq",
        "exl2",
        "exl3"
      ],
      "toolchain": "Python with a CUDA-capable GPU; installed via pip.",
      "blurb": "Consumer-GPU inference engine for EXL2 (and GPTQ) quantized checkpoints."
    },
    {
      "id": "ex-llama-v3",
      "displayName": "ExLlamaV3",
      "formats": [
        "exl2",
        "exl3"
      ],
      "toolchain": "Python with a CUDA-capable GPU; installed via pip.",
      "blurb": "Successor to ExLlamaV2 for the EXL3 quantization family."
    },
    {
      "id": "mlx",
      "displayName": "MLX",
      "formats": [
        "safetensors",
        "mlx"
      ],
      "toolchain": "Apple Silicon (M-series) on macOS; installed via pip (mlx-lm).",
      "blurb": "Apple's array framework — unified-memory inference for MLX-converted checkpoints."
    },
    {
      "id": "tensor-rt-llm",
      "displayName": "TensorRT-LLM",
      "formats": [
        "safetensors",
        "gptq",
        "awq",
        "tensor-rt-engine"
      ],
      "toolchain": "NVIDIA GPU with CUDA + TensorRT; engines are compiled per GPU architecture ahead of time.",
      "blurb": "NVIDIA's compiled-engine serving stack for maximum-throughput deployment."
    },
    {
      "id": "transformers",
      "displayName": "Transformers",
      "formats": [
        "safetensors",
        "gptq",
        "awq"
      ],
      "toolchain": "Python with PyTorch; installed via pip. CPU or GPU.",
      "blurb": "The Hugging Face reference runtime — broadest checkpoint coverage, not a serving-optimized path."
    },
    {
      "id": "ollama",
      "displayName": "Ollama",
      "formats": [
        "gguf",
        "remote"
      ],
      "toolchain": "The Ollama daemon installed locally; models are pulled by tag from the Ollama library on first use.",
      "blurb": "Daemon-managed local model runner — the provisioning path for every embedded catalog entry."
    },
    {
      "id": "cloud-serverless",
      "displayName": "Cloud serverless",
      "formats": [
        "remote"
      ],
      "toolchain": "No local toolchain. Network access plus a provider credential exposed as a ${ENV} variable.",
      "blurb": "A hosted HTTPS model API (see providers.json for WHO serves it)."
    },
    {
      "id": "external-cli",
      "displayName": "External CLI",
      "formats": [
        "remote"
      ],
      "toolchain": "The vendor's CLI binary installed and authenticated on the local machine (on PATH).",
      "blurb": "Delegates a turn to an installed vendor CLI instead of a raw model API."
    },
    {
      "id": "platform-funded",
      "displayName": "Platform funded",
      "formats": [
        "remote"
      ],
      "toolchain": "No local toolchain and no BYOK credential — execution is brokered by the OpenEng platform against account quota.",
      "blurb": "OpenEng-funded execution for accounts on a funded plan."
    }
  ],
  "compat": {
    "gguf": [
      "llama-cpp",
      "ollama",
      "vllm"
    ],
    "safetensors": [
      "vllm",
      "sglang",
      "tensor-rt-llm",
      "transformers",
      "mlx"
    ],
    "gptq": [
      "vllm",
      "tensor-rt-llm",
      "transformers",
      "ex-llama-v2"
    ],
    "awq": [
      "vllm",
      "tensor-rt-llm",
      "transformers"
    ],
    "exl2": [
      "ex-llama-v2",
      "ex-llama-v3"
    ],
    "exl3": [
      "ex-llama-v2",
      "ex-llama-v3"
    ],
    "mlx": [
      "mlx"
    ],
    "onnx": [
      "onnx-runtime"
    ],
    "tensor-rt-engine": [
      "tensor-rt-llm"
    ],
    "remote": [
      "cloud-serverless",
      "ollama",
      "external-cli",
      "platform-funded"
    ]
  },
  "limited": [
    [
      "gguf",
      "vllm"
    ]
  ]
}