{
  "model": "clef-vision-0.8b",
  "base_model": "Qwen/Qwen3.5-0.8B",
  "teacher": "Cloudflare/clef-flash",
  "contract": "Jev/SystemOne typed decisions over text + images: noul / choice / score, one logit per option",
  "packages": {
    "vision": {
      "dir": "Vision_P784",
      "file": "VisionTower_fp16.mlpackage",
      "patches": 784,
      "tokens": 196,
      "inputs": [
        "patches [784,1536]",
        "pos_embeds [784,768]",
        "cos [784,64]",
        "sin [784,64]",
        "mask [1,1,784,784]"
      ],
      "output": "tokens [196,1024]",
      "precision": "fp16 (default); see variants",
      "pos_embed": {
        "file": "Vision_P784/pos_embed.f32",
        "shape": [
          2304,
          768
        ],
        "dtype": "float32"
      },
      "variants": {
        "fp16": {
          "file": "VisionTower_fp16.mlpackage",
          "layout": "rows",
          "precision": "fp16",
          "note": "default: GPU, ~11 ms/image on M5 Pro; e2e prob drift 3.4e-3, argmax 100%"
        },
        "fp32": {
          "file": "VisionTower_fp32.mlpackage",
          "layout": "rows",
          "precision": "fp32",
          "note": "exact (7e-4 vs torch), GPU ~35 ms/image"
        },
        "ane_fp16": {
          "file": "VisionTowerANE_fp16.mlpackage",
          "layout": "channels_first",
          "precision": "fp16",
          "inputs": [
            "patches [1,1536,1,784]",
            "pos_embeds [1,768,1,784]",
            "cos [1,64,1,784]",
            "sin [1,64,1,784]",
            "mask [1,784(k),1,784(q)]"
          ],
          "output": "tokens [1,1024,1,196]",
          "note": "runs on the Neural Engine only with computeUnits=cpuAndNeuralEngine (ALL picks the GPU); ~24 ms/image, e2e prob drift 1.6e-2, argmax 100%"
        }
      },
      "default_variant": "fp16"
    },
    "lm": [
      {
        "dir": "LM_L512",
        "file": "LMRows_fp16.mlpackage",
        "length": 512,
        "inputs": [
          "hidden [1,512,1024]",
          "cos [512,64]",
          "sin [512,64]"
        ],
        "output": "states [1,512,1024]"
      },
      {
        "dir": "LM_L1024",
        "file": "LMRows_fp16.mlpackage",
        "length": 1024,
        "inputs": [
          "hidden [1,1024,1024]",
          "cos [1024,64]",
          "sin [1024,64]"
        ],
        "output": "states [1,1024,1024]"
      },
      {
        "dir": "LM_L2048",
        "file": "LMRows_fp16.mlpackage",
        "length": 2048,
        "inputs": [
          "hidden [1,2048,1024]",
          "cos [2048,64]",
          "sin [2048,64]"
        ],
        "output": "states [1,2048,1024]"
      }
    ],
    "head": [
      {
        "dir": "Head_L512_Q16_O64",
        "file": "Head_fp32.mlpackage",
        "length": 512,
        "max_questions": 16,
        "max_options": 64,
        "variants": {
          "fp32": {
            "file": "Head_fp32.mlpackage",
            "note": "GPU ~14 ms"
          },
          "fp16": {
            "file": "Head_fp16.mlpackage",
            "note": "ANE-eligible under ALL (~3.6 ms, 6 W ANE); e2e prob drift 3.6e-3, argmax 100%"
          }
        },
        "default_variant": "fp16"
      },
      {
        "dir": "Head_L1024_Q16_O64",
        "file": "Head_fp32.mlpackage",
        "length": 1024,
        "max_questions": 16,
        "max_options": 64,
        "variants": {
          "fp32": {
            "file": "Head_fp32.mlpackage",
            "note": "GPU ~14 ms"
          },
          "fp16": {
            "file": "Head_fp16.mlpackage",
            "note": "ANE-eligible under ALL (~3.6 ms, 6 W ANE); e2e prob drift 3.6e-3, argmax 100%"
          }
        },
        "default_variant": "fp16"
      },
      {
        "dir": "Head_L2048_Q16_O64",
        "file": "Head_fp32.mlpackage",
        "length": 2048,
        "max_questions": 16,
        "max_options": 64,
        "variants": {
          "fp32": {
            "file": "Head_fp32.mlpackage",
            "note": "GPU ~14 ms"
          },
          "fp16": {
            "file": "Head_fp16.mlpackage",
            "note": "ANE-eligible under ALL (~3.6 ms, 6 W ANE); e2e prob drift 3.6e-3, argmax 100%"
          }
        },
        "default_variant": "fp16"
      }
    ]
  },
  "embeddings": {
    "file": "embeddings.f16",
    "shape": [
      248320,
      1024
    ],
    "dtype": "float16",
    "tied_output_embeddings": true
  },
  "tokenizer": {
    "file": "tokenizer.json",
    "pad_token_id": 248044,
    "special_tokens": {
      "im_start": "<|im_start|>",
      "im_end": "<|im_end|>",
      "vision_start": "<|vision_start|>",
      "vision_end": "<|vision_end|>",
      "image_pad": "<|image_pad|>"
    }
  },
  "image_token_id": 248056,
  "vision_start_token_id": 248053,
  "vision_end_token_id": 248054,
  "image_processor": {
    "resample": "bicubic (torchvision, antialias)",
    "rescale_factor": 0.00392156862745098,
    "image_mean": [
      0.5,
      0.5,
      0.5
    ],
    "image_std": [
      0.5,
      0.5,
      0.5
    ],
    "patch_size": 16,
    "merge_size": 2,
    "temporal_patch_size": 2,
    "min_pixels": 16384,
    "max_pixels": 200704,
    "note": "min/max pixels are the training-time cap, not the processor defaults"
  },
  "vision_config": {
    "deepstack_visual_indexes": [],
    "depth": 12,
    "dtype": "bfloat16",
    "hidden_act": "gelu_pytorch_tanh",
    "hidden_size": 768,
    "in_channels": 3,
    "initializer_range": 0.02,
    "intermediate_size": 3072,
    "model_type": "qwen3_5_vision",
    "num_heads": 12,
    "num_position_embeddings": 2304,
    "out_hidden_size": 1024,
    "patch_size": 16,
    "rope_parameters": {
      "rope_theta": 10000.0,
      "rope_type": "axial"
    },
    "spatial_merge_size": 2,
    "temporal_patch_size": 2
  },
  "text_rope": {
    "mrope_interleaved": true,
    "mrope_section": [
      11,
      11,
      10
    ],
    "partial_rotary_factor": 0.25,
    "rope_theta": 10000000,
    "rope_type": "default"
  },
  "head_config": {
    "hidden_size": 1024,
    "width": 1024,
    "routing_layers": 2,
    "layers": 2,
    "heads": 16,
    "feedforward": 4096
  },
  "record_format": {
    "system_prompt": "Read the complete state and schema. Decide every field jointly. Each answer must be exactly one of that field's allowed options.",
    "source": "Cloudflare/clef joint_schema_model.py encode_record (Apache-2.0)"
  },
  "text_config_summary": {
    "hidden_size": 1024,
    "head_dim": 256,
    "num_hidden_layers": 24,
    "vocab_size": 248320,
    "rms_norm_eps": 1e-06
  },
  "compute_units_guidance": {
    "vision": "cpuAndGPU (fp16 rows) or cpuAndNeuralEngine (ane_fp16)",
    "lm": "cpuAndGPU (Qwen3.5 does not run on the ANE)",
    "head": "all (fp16 head lands on the ANE)"
  }
}