Laya for MLX-VLM

MLX-formatted weights for the original English/root checkpoint from convaiinnovations/laya, source revision 55cf4c4ebb4ebe31b2550e8bdf3bd21b99753851. This repository contains that variant only, not the multilingual or typed-decisions variants.

The weights retain their original precision and are not quantized. Weight names and packed QKV projections are converted to MLX-VLM's native layout. Encoder, decision-head, and calibration settings are included in root config.json, with tokenizer files at the root for standard loading.

Requires MLX-VLM with Laya support from PR #2397. Older releases without that support cannot load this checkpoint.

from mlx_vlm import load, predict

model, processor = load("nativ-community/laya")
result = predict(model, processor, "Please refund my duplicate charge.", {
    "department": {
        "type": "choice",
        "instructions": "Which team should handle this ticket?",
        "criteria": ["billing", "technical", "sales"],
    },
    "refund": {
        "type": "bool",
        "instructions": "Does this request ask for a refund?",
    },
})
print(result["answers"])

Supported question types are choice, score, and bool (noul is a Boolean alias). Scores use an ordered list of criteria. This is a decision model; it does not generate text.

Conversion verification

Strict loading succeeded before and after conversion. All 214 loaded tensors matched exactly in shape, dtype, and value. Nine answers across three short inputs and choice, Boolean, and score questions matched exactly, including probabilities, metadata, and token usage. This checks conversion preservation, not broad task accuracy or numerical identity with the upstream PyTorch runtime.

The original model is published under Apache-2.0. See the upstream model card for training details and intended use.