// the take
A one-person Core ML model zoo: PyTorch models from other open-source projects converted to .mlpackage, each with a conversion script and a SwiftUI demo app showing how to wire it up. It's for iOS/macOS developers who want a working on-device model (detection, segmentation, super-res, diffusion, TTS, small LLMs) without doing the PyTorch-to-CoreML conversion themselves.
Breadth is real and current — this isn't a stale 2021 zoo, it has YOLO26, RF-DETR, Qwen3.5, Gemma 4 and SAM2 alongside the older models, and it's being pulled into real projects (xLights downloads the HTDemucs model directly for stem separation). Conversion scripts ship alongside the weights, so you're not stuck with a black box if you need to tweak input size or quantization. Sample apps show actual integration work most people get wrong on the first pass — NMS, cxcywh decoding, BPE tokenizers, Kalman-filter tracking — not just `model.predict()`.
Weights are distributed as Google Drive links and GitHub release zips, not through any package manager — no checksums, no guaranteed uptime, and the older Drive links are the kind of thing that quietly 404s in a year. License is 'whatever the original project used,' so a single repo mixes MIT, Apache, GPL-3.0 and AGPL-3.0 model weights — ship one of the YOLOv9/v10/11/26/World/E or FastSAM models in a closed-source App Store app and you've likely pulled in AGPL/GPL obligations without realizing it. It's maintained by one person across an enormous surface area (vision, audio, diffusion, LLMs); there's no CI checking that old .mlpackage downloads still load against the current Core ML runtime, so models added years ago carry real bit-rot risk. No versioning or migration notes per model, so an app already shipping an older YOLO or SAM variant has no signal when a newer, better one replaces it in the table.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →