z0lo/LFM2.5-VL-1.6B-Extract-4bit

LFM2.5-VL-1.6B-Extract — MLX 4-bit

This repository contains an MLX affine 4-bit conversion of LiquidAI/LFM2.5-VL-1.6B-Extract at the immutable upstream revision 21073aa65b326c125ceb07d80fa08900bc2eb0d9.

This is the Extract checkpoint, not the general LiquidAI/LFM2.5-VL-1.6B checkpoint and not mlx-community/LFM2.5-VL-1.6B-4bit.

Artifact details

Property	Value
Architecture	LFM2.5-VL: 1.2B LFM2 language model plus approximately 400M SigLIP2 vision encoder
Model type	lfm2_vl
Weight format	MLX-compatible safetensors
Quantization	MLX affine 4-bit
Quantization group size	64
Evaluated conversion runtime	mlx-vlm==0.6.7
Model weights	1,511,170,678 bytes
Complete runtime snapshot	1,515,976,304 bytes

This is an MLX checkpoint. It is not a Core ML model and is not the original upstream BF16 Transformers checkpoint. A runtime must explicitly support the lfm2_vl architecture; the presence of MLX weights does not imply compatibility with every MLX or MLX Swift release.

Important configuration note

The converted config.json intentionally sets:

{
  "text_config": {
    "block_ff_dim": 12288
  }
}

The upstream source configuration omits this field. During conversion, mlx-vlm==0.6.7 otherwise selected the 450M-model default and produced incorrect feed-forward tensor shapes. The value 12288 matches the Extract checkpoint's source tensors. No source tensors were changed or dropped.

Loading with MLX-VLM

The conversion was evaluated with:

pip install mlx-vlm==0.6.7

from mlx_vlm import load

model, processor = load(
    "z0lo/LFM2.5-VL-1.6B-Extract-4bit",
    adapter_path=None,
    lazy=False,
    trust_remote_code=True,
)

Use deterministic greedy decoding (temperature=0) for extraction, as recommended by the upstream model card.

Intended use and limitations

The upstream model is designed for a single image, a YAML field list in the system prompt, and a flat JSON object as output. This conversion was created for on-device structured-extraction evaluation.

Quantization and runtime differences can affect model quality. JSON or schema validity does not establish factual accuracy. Validate extracted values against the source image before applying them to production data.

This converted checkpoint has not been independently qualified for arbitrary multi-image reasoning, free-form visual question answering, or every MLX runtime.

License and attribution

The original model and this converted checkpoint are governed by the LFM Open License v1.0. See LICENSE and review its terms before using or redistributing the model.

Original model: Copyright Liquid AI, Inc.