laya mlx

Overview

laya mlx is a native MLX FP16 conversion of convaiinnovations/laya for Apple silicon. It provides a bidirectional decision encoder with Laya’s decision Transformer, scoring head, and action head. The model is hosted on Hugging Face under the Apache-2.0 license.

Key Features

• Loads through the dedicated laya-mlx runtime and runs model computation in MLX.
• Supports choice, ordinal score, and boolean noul question types.
• Uses ModernBERT-large with a 512-token total context.
• Preserves FP16 weights and supports dtype=float32 for closer agreement with upstream FP32 arithmetic.
• Does not require PyTorch or Transformers; it is not a generative language model and does not include training code.

How It Works

1. Install laya-mlx on an Apple silicon Mac running macOS 14+ and Python 3.11+.
2. Load aac6fef/laya-mlx with laya.load(...).
3. Pass a text request and a dictionary of typed questions with instructions and options to agent.predict(...).
4. Read the structured answers from result["answers"].

The model card example routes a duplicate-billing request to a department choice and checks whether the customer asks for a refund.

Validation and Limits

The maintainer reports local tests on an Apple M3 Max with MLX 0.32.2, including 63/63 argmax agreement across 16 cases, a maximum calibrated probability difference of 0.0054443, deterministic outputs across 100 repeated calls, and exact FP16 tensor checks. These measurements document port fidelity, not universal answer correctness. The checkpoint is specialized for upstream workflows; language and task limitations remain those of the original model. The model card notes that a multilingual checkpoint is intended for non-English text.

Pricing and Availability

No paid plan or hosted inference offering is published. The model is downloadable from Hugging Face and is tagged MLX, Safetensors, English, Apple silicon, and decision-model. The page says it is not currently deployed by an Inference Provider.

Screenshots