Laya Windows

The open-source, self-hosted decision engine for Windows

Typed, non-generative AI decisions — choice / score / noul — in a single forward pass, with calibrated probabilities and zero hallucination risk. Ported off macOS-only Core ML onto ONNX Runtime + DirectML, so it runs on any Windows PC.

Same category as TypeSafe's Jev — open-weight, Apache-2.0, and yours to run.

🌐 العربية · 中文 · Español · Português · हिन्दी · Русский · 日本語 · Deutsch · Français · 한국어

---

TL;DR: laya-coreml only runs on Apple Silicon Macs. Laya Windows ports the same Laya typed-decision model (no text generation, just calibrated choice/score/noul probabilities) onto ONNX Runtime + DirectML, so it runs on any Windows PC with any DX12 GPU, with a dedicated Arabic evaluation suite the original doesn't have.

Why this exists

laya-coreml ports the Laya typed-decision model onto Apple's Neural Engine. macOS and Apple Silicon only. Most of the world, and most of our own stack, runs Windows. Laya Windows ports the same idea onto ONNX Runtime + DirectML, Microsoft's cross-vendor GPU inference stack, so it runs on any DirectX 12 GPU (NVIDIA, AMD, Intel, integrated) with no CUDA requirement, plus a first-class Arabic demo and evaluation suite the original project doesn't ship.

Laya itself evaluates typed questions (choice, score, noul) over any text or JSON state in a single forward pass. No autoregressive decoding, nothing to parse. You get calibrated probabilities back, not a wall of generated text.

# Full typed-decision API (CPU/CUDA via PyTorch — works today, any OS):
import laya
agent = laya.load("convaiinnovations/laya-multilingual")
result = agent.predict(
    "تم خصم المبلغ مرتين من حسابي، الرجاء استرجاع الفرق بأسرع وقت",
    {
        "refund": {"type": "noul", "instructions": "Does the customer request a refund?"},
        "urgency": {"type": "score", "criteria": ["not urgent", "soon", "critical"]},
    },
)
print(result["answers"]["refund"])

# Windows-native fast path (ONNX + DirectML encoder — see Status below):
from laya_windows import LayaWindowsEngine
engine = LayaWindowsEngine.from_dir("models")
vector = engine.embed("تم خصم المبلغ مرتين من حسابي")

Laya Windows vs. TypeSafe Jev

Jev is a commercial "System One" model from TypeSafe, announced in September 2026, in the same non-generative, typed-probability category as Laya (choice/ score/noul answers, no token generation). It's closed-weight and API-metered. Laya Windows is the open-weight route to the same category: self-hosted, Apache-2.0, no per-call billing, running on your own Windows GPU instead of TypeSafe's cloud.

	TypeSafe Jev	Laya Windows
Weights	Closed, API-only	Open (Apache-2.0), self-hosted
Cost model	Per-token API pricing	Free, your own hardware
Deployment	Cloud API	Local — Windows, any DX12 GPU or CPU
Data	Leaves your machine	Stays local
Arabic	Not documented	Dedicated benchmark suite

What's different from the Mac-only original

	laya-coreml (upstream)	Laya Windows
Platform	macOS 15+, Apple Silicon only	Windows 10/11, any DX12 GPU
Accelerator	Core ML / Apple Neural Engine	ONNX Runtime / DirectML (GPU-agnostic)
Quantization	Core ML W8 palette	ONNX dynamic int8
Arabic support	Inherited from base model, untested	First-class: dedicated eval suite + demo below
Evaluation	Public benchmark fixtures	Same fixtures + an Arabic/English triage benchmark against real support-message phrasing

Benchmarks

Live, searchable results: zuhair-01.github.io/laya-windows

A sortable table, generated only from real committed runs of bench/run_eval.py (department routing, urgency scoring, refund-intent detection, Arabic + English mixed). No estimated numbers are published here or in this README; the page says so plainly when it's empty. See bench/ for the harness itself, and CONTRIBUTING.md to add your own hardware's results.

Install

git clone https://github.com/Zuhair-01/laya-windows
cd laya-windows
pip install -r requirements.txt

# 1. Export the encoder to ONNX (one-time, per checkpoint)
python onnx_export.py --model convaiinnovations/laya-multilingual --out models/laya-multilingual.onnx

# 2. Optional: int8 quantize for lower latency / smaller footprint
python quantize.py --in models/laya-multilingual.onnx --out models/laya-multilingual.int8.onnx

# 3. Run the Arabic triage demo
python demo/triage_demo.py

Requires Windows 10 1903+ / Windows 11, Python 3.10+. DirectML ships in the box on modern Windows — no separate GPU driver stack to install beyond your normal GPU driver.

Status

• Encoder export (PyTorch → ONNX)
• DirectML / CPU inference engine
• Dynamic int8 quantization
• Arabic + English triage benchmark harness
• Decision-head (RL projection layer) full ONNX port. Currently loaded from the original transformers checkpoint; encoder runs on DirectML, head runs on CPU. Tracked in #1.

This project is upfront about what's ported and what isn't — same spirit as upstream's own "limits" section. No inflated claims.

Use cases

• Support/lead triage: route incoming messages by department, urgency, refund/churn intent, in Arabic or English, without an LLM call.
• Local classification gate: cheap first-pass filter (spam, urgency, intent) ahead of an LLM, cutting API cost on high-volume inboxes.
• Offline/on-prem deployments: no cloud dependency, runs on a normal Windows machine.

FAQ

Does this need a Mac or an Apple Neural Engine? No. That's the whole point — it runs on plain Windows 10/11 with any DirectX 12 GPU (NVIDIA, AMD, Intel, or integrated), via ONNX Runtime + DirectML instead of Core ML.

Is this the same model as laya-coreml? Yes, same underlying Laya weights (convaiinnovations/laya-multilingual etc. on Hugging Face) — different execution backend. Accuracy is identical; only latency and platform differ.

Does it generate text like an LLM? No. It answers typed questions (choice, score, noul) with calibrated probabilities in a single forward pass. There's no decoding step and nothing to hallucinate.

Is Arabic actually supported, or just "should work"? The laya-multilingual checkpoint is trained on 100+ languages including Arabic. This repo adds a dedicated Arabic/English triage benchmark (see bench/) to verify it, rather than assuming multilingual claims transfer without checking.

Can I run this without a GPU? Yes — it falls back to CPUExecutionProvider automatically if DirectML isn't available.

Credits

Independent Windows port of Laya by Convai Innovations, following the same approach as the Apple-only laya-coreml port. Not an official Convai Innovations, Apple, or Microsoft release. Apache-2.0, see NOTICE.

---

_{Keywords: Laya Windows, Core ML alternative Windows, ONNX Runtime DirectML NLP, typed decision model, Arabic text classification offline, local LLM alternative Windows, run Apple Neural Engine model on Windows, Arabic intent classification, offline support ticket triage, GPU-agnostic transformer inference.}