User repository on GitHub
Correctness-validated heterogeneous Laya runtime for Apple Silicon (MLX GPU + Apple Neural Engine)
- Forks
- 8
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
Laya Apple
A correctness-validated Laya runtime for Apple Silicon that runs upstream Laya on the MLX GPU and the Apple Neural Engine, using the ANE only after an artifact passes a parity gate on your machine, plus a local Jev-compatible server.
- Stars
- 19
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 3 min
- Published by
- Community
- Author
- tc3oliver
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceRuns locally
Highlights
- Runs upstream Laya on the MLX GPU and the Apple Neural Engine, picking one engine per request based on question length and artifact validation.
- A plain Core ML export ran error-free on the ANE yet disagreed with upstream on up to 85 decisions, so an ANE artifact is used only after a parity gate.
- Each result records its device and routing_reason; asking for the ANE explicitly returns the validated artifact or an exception, never a silent fallback.
- With execution="workers" both engines serve at once; measured on one Apple M4 Max and watchable with uvx laya-apple switchyard.
- The serve extra makes laya-apple serve answer existing Jev clients locally with upstream Laya, with no changes to the client.
Quickstart
bash
pip install 'laya-apple[ane]'Watch out
Grounded in the truncated repository README and GitHub metadata reviewed on 2026-10-15; the speed and routing numbers come from one Apple M4 Max, the ANE extra needs macOS 26 or 27 with coremltools 9.0, and the README's build-from-source section was cut off.
Questions about this entry
Short answers drawn from the linked project or page.
It runs upstream Laya on both the MLX GPU and the ANE and picks one engine per request. Single-question requests go to the ANE only when a validated artifact is present; long and multi-question requests stay on the GPU.
The README reports that a plain Core ML export ran on the ANE without a single error and still disagreed with upstream on up to 85 decisions, so laya-apple runs its parity gate on your machine and only then trusts the ANE path.
On Apple silicon with Python 3.11 to 3.13, run pip install 'laya-apple[ane]' for the MLX GPU plus Neural Engine runtime. Without the ane extra, or without a built ANE artifact, everything runs on the MLX GPU.
Yes, with the serve extra: laya-apple serve answers existing Jev clients locally using upstream Laya, and the README says no client changes are needed.
Every result records which device answered and a routing_reason string. The README's example output shows ane validated_short_single_question_path 11.2 ms; without an ANE artifact the reason explains why the request ran on MLX instead.
No key is involved: the first call downloads the pinned checkpoint, and after that it works offline with local_files_only=True. The ANE artifact can be fetched prebuilt on a supported Mac or built locally with the convert extra.
It is tc3oliver's project under Apache-2.0, and the ANE path needs Apple silicon with macOS 26 or 27 and coremltools 9.0. The README's performance figures come from a single Apple M4 Max, so treat them as one machine's results.
More Jev resources like this
An independent Apple Silicon runtime for Convai Innovations' Laya, a 421M ModernBERT-large typed-decision model: MLX GPU inference with optional Core ML Neural Engine bodies and a measured ANE/GPU batch router.
Jev Repos & SDKs#community#catalog-import#madewithjev
Community
djlougen/laya-fastPython
djlougen/laya-fastPythonLaya GGUF inference and an HTTP server written in Goish Rust: one Linux binary runs upstream Laya or OpenThai-SystemOne on your CPU and serves typed choose, score and check decisions over HTTP, the CLI or a JSON-lines process.
Jev Repos & SDKs#community#catalog-import#madewithjev
Community
cogentica-ai/laya-goishRust
cogentica-ai/laya-goishRustAn Ollama-style local runner for System One decision models from Hugging Face: pull, serve and query them behind the same /v1/systemone API as TypeSafe's Jev, with a terminal model manager and a browser playground.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
nvkudva/ollajevPython
nvkudva/ollajevPythonA 100% offline, zero-cost Rust drop-in for the TypeSafe Jev System One API: one native binary serves Choice, Score, Noul and Route decisions on CPU in under 18 ms with no API keys, cloud egress or rate limits.
Jev Repos & SDKs#community#catalog-import#madewithjev
Community
ronitgupta138/local-jevRust
ronitgupta138/local-jevRustA lightweight Python engine that mimics Jev's structured decisions on local LLMs via Ollama, vLLM or llama.cpp: choice, score and noul questions come back as JSON with probabilities and confidence, plus an OpenAI-compatible gateway.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
tapsin/jev-localPython
tapsin/jev-localPythonA Jev-kind model inference engine and compatible API server in Python: each candidate path runs as segmented forwards with prefix reuse, and the Jev wire contract is served on top with NanoJev as the wired backend. The README reports 2.57x single-request throughput on MPS at 100% argmax agreement.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
zerodegress/jevinfPython
zerodegress/jevinfPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.