{
  "schema_version": 1,
  "model_id": "EldanRing/Winnow-E4B",
  "base_model": "google/gemma-4-E4B-it",
  "base_revision": "ee0ef6023621cff504d758262d4e04895a5af4a2",
  "license": "apache-2.0",
  "training_method": "rank-32 alpha-64 LoRA on language tensors; FP32 adapter merge",
  "converter_revision": "911f6cdc8ab8a530b2bee09ee61471a6f3178eeb",
  "inference_source_commit": "77d1458",
  "artifacts": {
    "q8_0": {
      "path": "gguf/Winnow-E4B-Q8_0.gguf",
      "bytes": 8005437472,
      "sha256": "840e3f50e5a9c218727f44e121d1b37cc9e2c3b318c8eb422ba6ef2e27b618a2"
    },
    "bf16": {
      "path": "gguf/Winnow-E4B-BF16.gguf",
      "bytes": 15053096992,
      "sha256": "53c3e504c061f6a7ce84cb6240d427ac6c630850a2cf93f3c2ee980c8136e7e7"
    },
    "projector_f16": {
      "path": "gguf/mmproj-Winnow-E4B.gguf",
      "bytes": 990372672,
      "sha256": "ddf46c21d7078e95338cfc22306b19b276a29a5ad089023449dd54d4b6170a51"
    },
    "mtp_assistant_bf16": {
      "path": "gguf/Gemma-4-E4B-IT-Assistant-BF16.gguf",
      "sha256": "4e3c9d335b248ced9bd3e0584547efca30dbf6cfd0debb810613f811a575162a",
      "bytes": 171766688,
      "source_model": "google/gemma-4-E4B-it-assistant",
      "source_revision": "8d0031ea8c2109e2b1e86bb9368a4539b537f80a",
      "license": "apache-2.0",
      "documentation": "docs/assistants/README.md"
    }
  },
  "decision_calibration": {
    "artifact": "q8_0",
    "temperature": 1.2574172017327816,
    "calibration_rows": 778,
    "scope": "text decisions; 8K-context deployment"
  },
  "bf16_decision_calibration": {
    "artifact": "bf16",
    "temperature": 1.3331553765162731,
    "calibration_rows": 778,
    "scope": "text decisions; 8K-context deployment"
  },
  "measured_profile": {
    "gpu": "NVIDIA RTX 5070 Ti",
    "context": 8192,
    "kv_cache": "q8_0",
    "decision_parallel": 4,
    "modality": "text"
  },
  "recommended_q8_gpu_profiles": {
    "text_8k": {
      "recommended_vram_gb": 10,
      "context": 8192,
      "projector": null,
      "observed_peak_memory_mib": 8769
    },
    "vision_64k": {
      "recommended_vram_gb": 12,
      "context": 65536,
      "projector": "projector_f16",
      "observed_peak_memory_mib": 10979
    },
    "basis": "Capacity recommendations from RTX 5070 Ti measurements; not qualification of every GPU. Q8 KV, four decision branches, one chat slot, exclusive memory."
  },
  "vision_64k_smoke_profile": {
    "artifact": "q8_0",
    "projector": "projector_f16",
    "gpu": "NVIDIA RTX 5070 Ti",
    "context": 65536,
    "kv_cache": "q8_0",
    "decision_parallel": 4,
    "scope": "synthetic-image operational smoke test; not broad vision or long-context quality"
  },
  "training_pipeline_released": false,
  "optional_inference": {
    "repository": "https://github.com/EldanRing/winnow-inference",
    "runtime_lock_sha256": "9560b74c9fd736c80880c2733569075e066e7e8f49a7b465c8a4589604c4f65e",
    "context": 8192,
    "native_parallel": 4,
    "chat_parallel": 1,
    "cache": "q8_0",
    "scope": "Linux/CUDA singleGPU tested RTX5070Ti16GB; experimental adaptive calibration text only",
    "supported_presets": [
      "e4b-q8-vision8k-mtp"
    ],
    "adaptive_policies": [
      "e4b-calibrated50-v1"
    ]
  }
}