{
  "artifacts": [
    {
      "file": "LFM2.5-Encoder-230M_wi8fc.tflite",
      "sha256": "6be33404190793ec1bee8b054bb0903e2bf5a33b266aa224a96c74a829247743",
      "size_mb": 234.397
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "python scripts/convert_lfm25_encoder.py LiquidAI/LFM2.5-Encoder-230M out/lfm25_encoder_230m",
    "quantization": "int8 dynamic-range (linears + embedding; convs float) — wi8fc variant",
    "tool": "litert-torch (litertlm-convert scripts/convert_lfm25_encoder.py, litert_torch.signature multi-signature trace)",
    "tool_version": "0.9.2 (stock release; .venv-092 per lfm25_encoder_work/FINDINGS.md)"
  },
  "cross_runtime": [],
  "delegation": null,
  "model": {
    "family": "lfm2.5",
    "id": "lfm25-encoder-230m",
    "license": "lfm1.0",
    "source_url": "https://huggingface.co/LiquidAI/LFM2.5-Encoder-230M",
    "task": "feature-extraction"
  },
  "pitfalls": [
    "The fp16 variant is desktop-oriented: XNNPACK's per-signature fp32 unpacking is heavy on phone memory limits — use the int8 (wi8fc) file on mobile (iPhone-verified bit-exact).",
    "All signatures are batch-1, right-padded static shapes (encode_64/128/256/512, mlm_128); padded positions are fully masked in-graph, so outputs at valid positions are independent of padding length.",
    "iPhone 17 Pro int8 run reproduces Mac outputs bit-exactly, but peak footprint is ~1.0 GiB.",
    "License: LFM Open License v1.0 — note the commercial-use threshold (Section 5); redistributed as Derivative Works with modification notices per Section 4."
  ],
  "schema_version": "1.0"
}