OpenJev Models: Core ML packages for iOS and macOS

Core ML model packages used by OpenJevSwift for typed AI decisions on iOS and macOS (iPhone and Mac). Nothing here is trained by this organisation: each package is a conversion of another author's checkpoint, published under that checkpoint's license. The weights are the authors' own, stored as float16 in Apple's Core ML format.

You do not need to download anything by hand. OpenJevSwift's OpenJevEncoders library, and the openjev server built on it, download the package a backend needs on first use, check every file's SHA-256 against the digests compiled into the library, and keep the package in ~/Library/Application Support/OpenJevSwift/encoders. To run without downloads, set OPENJEV_ENCODER_MODELS to a folder of converted packages, as OpenJevSwift's docs/deployment.md describes.

Each release holds one package version. Its assets are the files of the .mlpackage, uploaded under their paths with / replaced by --, so a device downloads them one by one with no archive to unpack. How the packages are converted, checked and published is documented in OpenJevSwift under Tools/encoders/ and decision D-033.

Core ML model packages

Release	Model	Runs on	Input shapes	Size
verdict-m18-fp16-v1	verdict-1.4, Verdict by Heman10x	Mac (GPU) and iPhone (Neural Engine)	one function per shape: 1 or 16 questions by 128, 256 or 512 tokens	306 MB
laya-m18-fp16-v1	laya-1.0, Laya by Nandakishor M / Convai Innovations	Mac (GPU)	one function per shape: 1 or 16 questions by 128, 256, 512 or 1,024 tokens	849 MB
laya-f18-b1s128-fp16-v1	laya-1.0	iPhone (Neural Engine)	one question by 128 tokens	843 MB
laya-f18-b1s256-fp16-v1	laya-1.0	iPhone (Neural Engine)	one question by 256 tokens	843 MB
laya-f18-b1s512-fp16-v1	laya-1.0	iPhone (Neural Engine)	one question by 512 tokens	843 MB
laya-f18-b1s1024-fp16-v1	laya-1.0	iPhone (Neural Engine)	one question by 1,024 tokens	845 MB

An iPhone app fetches only the Laya lengths it asks for (LayaBackend.prefetch(lengths:)), at least the 128-token package before its first read.

A release name reads as model, layout and target, weights, then version:

• m is a multifunction package, one function per input shape; f is a fixed-shape package holding one program (b1s128: batch 1, that is one question, by 128 tokens).
• 18 is the Core ML target, iOS 18 (macOS 15 on a Mac).
• fp16 is float16 weights.
• v1 is the first release of that package. A changed package gets a new release; an existing one is never replaced, since OpenJevSwift pins every file's digest.

Credits

• Verdict is by Heman10x (Verdict-open-jev): a ModernBERT-base encoder with a GLiClass head, calibrated with RLCD. The package here is a float16 Core ML conversion of the author's checkpoint, heman10x/rlcd-modernbert-151m at 8af2496 (Apache-2.0), with one function per input shape for iOS 18 and macOS 15. The author's tokenizer and calibration files are not re-hosted; OpenJevSwift reads them from the checkpoint on Hugging Face at the pinned revision.
• Laya is by Nandakishor M / Convai Innovations (laya): a ModernBERT-large model (421M parameters) fine-tuned on the typed-decisions workflows. The packages here are float16 Core ML conversions of the author's checkpoint, convaiinnovations/laya-typed-decisions at 1a793eb (Apache-2.0): laya-m18-fp16 for a Mac's GPU, and four fixed-shape packages for the iPhone's Neural Engine, which does not load Laya's multifunction package. The author's tokenizer and rl_agent_config.json are not re-hosted; OpenJevSwift reads them from the checkpoint on Hugging Face at the pinned revision.

The packages are provided as is, under the Apache License 2.0 of the checkpoints they come from.