{
  "schema_version": 1,
  "generated_at": "2026-10-04T13:50:01.668256Z",
  "models": [
    {
      "name": "clef",
      "aliases": [],
      "family": "qwen35",
      "min_shiso_version": "0.1.0",
      "description": "Cloudflare Clef decision model (Qwen3.8-27B backbone with a joint schema head) for /v1/systemone and /v1/decisions",
      "tags": [
        "decision",
        "systemone"
      ],
      "tasks": [
        "decision"
      ],
      "inputs": [
        "text"
      ],
      "outputs": [
        "text"
      ],
      "updated_at": "2026-10-04T16:50:01+03:00",
      "manifest_sha256": "5eb498590219953b1202caf48953da238edc24d3837b53c34d24f289fbe3c332",
      "variants": [
        {
          "key": "q4_k_m",
          "backends": [
            "cuda",
            "metal"
          ],
          "quant": "Q4_K_M",
          "min_ram_bytes": 34359738368,
          "disk_bytes": 19252222358,
          "min_cuda_sm": 80
        }
      ]
    },
    {
      "name": "clef-flash",
      "aliases": [],
      "family": "qwen35",
      "min_shiso_version": "0.1.0",
      "description": "Cloudflare Clef-Flash decision model (Qwen3.5-9B backbone with a joint schema head) for /v1/systemone and /v1/decisions",
      "tags": [
        "decision",
        "systemone"
      ],
      "tasks": [
        "decision"
      ],
      "inputs": [
        "text"
      ],
      "outputs": [
        "text"
      ],
      "updated_at": "2026-10-04T16:50:01+03:00",
      "manifest_sha256": "0d258ef29ac191f405eb34fb1bded92403354840a7e41a7ab8d6d8d3238c373e",
      "variants": [
        {
          "key": "q4_k_m",
          "backends": [
            "cuda",
            "metal"
          ],
          "quant": "Q4_K_M",
          "min_ram_bytes": 8589934592,
          "disk_bytes": 6506449296,
          "min_cuda_sm": 80
        }
      ]
    },
    {
      "name": "flux.2-klein-4b",
      "aliases": [],
      "family": "flux2",
      "min_shiso_version": "0.1.0",
      "description": "FLUX.2 [klein] 4B step-distilled text-to-image model, FP8 transformer with a Qwen3 text encoder",
      "tags": [
        "image-generation"
      ],
      "tasks": [
        "image_generation"
      ],
      "inputs": [
        "text"
      ],
      "outputs": [
        "image"
      ],
      "updated_at": "2026-10-04T14:53:43+03:00",
      "manifest_sha256": "bff1169737dedf542737ffd5b6ff80baf7843396e5ae202115e10f2b334ad067",
      "variants": [
        {
          "key": "fp8",
          "backends": [
            "metal",
            "cuda"
          ],
          "quant": "fp8",
          "min_ram_bytes": 17179869184,
          "disk_bytes": 12299645375
        }
      ]
    },
    {
      "name": "gemma-4-e4b",
      "aliases": [],
      "family": "gemma4",
      "min_shiso_version": "0.1.0",
      "description": "Gemma 4 E4B instruction-tuned, quantization-aware trained Q4_0, with vision and audio towers",
      "tags": [
        "vision",
        "audio",
        "tools"
      ],
      "tasks": [
        "chat"
      ],
      "inputs": [
        "text",
        "image",
        "audio"
      ],
      "outputs": [
        "text"
      ],
      "updated_at": "2026-10-04T15:40:24+03:00",
      "manifest_sha256": "6637b873cddcb0b7ff819567dddd1e75dbffcc964b103c10acab7f5757038726",
      "variants": [
        {
          "key": "qat-q4_0",
          "backends": [
            "cuda",
            "metal"
          ],
          "quant": "Q4_0 (QAT)",
          "min_ram_bytes": 8589934592,
          "disk_bytes": 6178692489,
          "min_cuda_sm": 80
        }
      ]
    },
    {
      "name": "qwen3.5-4b",
      "aliases": [],
      "family": "qwen35",
      "min_shiso_version": "0.1.0",
      "description": "Dense 4B hybrid-attention model with a vision tower",
      "tags": [
        "vision",
        "tools"
      ],
      "tasks": [
        "chat"
      ],
      "inputs": [
        "text",
        "image",
        "video"
      ],
      "outputs": [
        "text"
      ],
      "updated_at": "2026-10-04T10:41:33+03:00",
      "manifest_sha256": "ffb670e004ee5d21d962e934bf02bf3230721e8df996135322e9f7e3ba96d6f6",
      "variants": [
        {
          "key": "mlx-optiq-4bit",
          "backends": [
            "metal"
          ],
          "quant": "mlx-optiq-4bit",
          "min_ram_bytes": 8589934592,
          "disk_bytes": 3956912147
        },
        {
          "key": "q4_k_m",
          "backends": [
            "cuda",
            "metal"
          ],
          "quant": "Q4_K_M",
          "min_ram_bytes": 8589934592,
          "disk_bytes": 3426197888,
          "min_cuda_sm": 80
        }
      ]
    },
    {
      "name": "qwen3.8-27b",
      "aliases": [],
      "family": "qwen35",
      "min_shiso_version": "0.1.0",
      "description": "Dense 27B hybrid-attention model with a DFlash2 block-diffusion drafter",
      "tags": [
        "tools"
      ],
      "tasks": [
        "chat"
      ],
      "inputs": [
        "text"
      ],
      "outputs": [
        "text"
      ],
      "updated_at": "2026-10-04T10:59:32+03:00",
      "manifest_sha256": "f474adf277fc0f00ed87d131d3ec6c9fbf097bfaf3618edcd1185183ec0e7980",
      "variants": [
        {
          "key": "mlx-4bit",
          "backends": [
            "metal"
          ],
          "quant": "mlx-4bit",
          "min_ram_bytes": 34359738368,
          "disk_bytes": 19930307866
        },
        {
          "key": "nvfp4",
          "backends": [
            "cuda"
          ],
          "quant": "nvfp4",
          "min_ram_bytes": 34359738368,
          "disk_bytes": 25793770138,
          "min_cuda_sm": 80
        }
      ]
    },
    {
      "name": "qwen3.8-flash",
      "aliases": [
        "qwen3.8-flash-next"
      ],
      "family": "qwen4exp",
      "min_shiso_version": "0.1.0",
      "description": "Sparse hybrid-attention MoE with n-gram embeddings, built-in MTP draft head and vision tower",
      "tags": [],
      "tasks": [
        "chat"
      ],
      "inputs": [
        "text",
        "image",
        "video"
      ],
      "outputs": [
        "text"
      ],
      "updated_at": "2026-10-04T01:43:15+03:00",
      "manifest_sha256": "22f9d4d22ba2b7f532b663ea1d16dd6123247b2fd6d1913a440e66410fe39cc7",
      "variants": [
        {
          "key": "exl3-405",
          "backends": [
            "cuda"
          ],
          "quant": "exl3-4.05bpw",
          "min_ram_bytes": 68719476736,
          "disk_bytes": 107462993908
        }
      ]
    }
  ]
}