Model catalog

For LYNX SDK 1.0. Single source of truth for model facts — recipes and the lynx_find_model tool read from here. Load any model by its slug: Lynx.load("<slug>") (Swift) / lynx.open("<slug>") (Python) / Lynx.load("<slug>") (Kotlin/Java) / lynx_open("<slug>", …) (C).

Two models are available. Both are keyless (no API key — the SDK mints a per-device trial on first load), production version 2.0, and ship for linux-x64 (ONNX), ios-arm64 (CoreML), android-arm64-v8a (TFLite), input 640×640.

slugtaskclassesheadsuse it for
lynx-basicgeneral object detection80 (COCO 2017: person, bicycle, car, … toothbrush)detection, segmentation, pose (keypoints), depth"detect / segment / find objects / poses / depth" on everyday scenes
lynx-ocr-fleettext / OCR (character detection)67 character classes — 0–9, A–Z, a–z, and - / . ( )detection (one detection per character)reading short text / IDs / codes off an image

Class 0 (qr, a reserved anti-tamper channel) is hidden by the SDK — model.classes and detection class ids are the public set above (80 for basic, 67 for ocr-fleet).

How to choose

  • Detect everyday objects → lynx-basic. Detection uses the box head every model has; this model also exposes segmentation, pose, depth on the same Frame (realize per-detection ROI heads, or read frame.depth). Confirm at runtime with model.capabilities.
  • Read text / characters / a code or ID → lynx-ocr-fleet. It detects individual characters as boxes; you order them left-to-right and read each detection's class name (the class name is the character). See the OCR recipe.
  • A class not in COCO-80, or oriented/rotated objects, or any task neither model covers → no off-the-shelf fit: go to no-model.md. That's a real path (synthetic data → trained .lnx → same API), not a dead end.

Loading specifics

  • Both are keyless: Lynx.load("lynx-basic") / lynx.open("lynx-ocr-fleet") just works — no setApiKey.
  • First load downloads + verifies + caches (a few seconds); subsequent loads are local. Run the first load off the main thread.
  • Pin a build with version:/version= (e.g. "2.0"); omit for the current production version.

View this page as raw Markdown — the exact content coding agents fetch. Index: /llms.txt.