---
library_name: transformers
license: other
license_name: lfm1.0
license_link: LICENSE
pipeline_tag: image-text-to-text
base_model: LiquidAI/LFM2.5-VL-450M
tags:
- lfm2.5-vl
- no-tiling
---

# LFM2.5-VL-450M — no tiling

**Modification notice:** Modified by samuelstevens: image_processor.do_image_splitting changed from true to false, based on LiquidAI/LFM2.5-VL-450M@fc6221ca597f3315e4f82fc2df606783267b34ba. Model weights are unchanged. This is a processor-configuration fork, not a fine-tuned model.

This repository contains the complete original checkpoint from
[LiquidAI/LFM2.5-VL-450M](https://huggingface.co/LiquidAI/LFM2.5-VL-450M/tree/fc6221ca597f3315e4f82fc2df606783267b34ba),
with its original weights, tokenizer, chat template, model configuration and
license. The only behavioral configuration change is:

```json
{"image_processor": {"do_image_splitting": false}}
```

A modification notice was also added to `processor_config.json` as metadata.
The image processor uses a single resized whole-image view instead of tiles.
Other image-processing defaults are unchanged, including `max_image_tokens=256`.
This trades visual detail for lower image-token usage; no accuracy improvement
or equivalence is claimed. No SFT, model inference, or benchmark was performed
as part of creating this fork. No FishVista images or annotations are included.

## Usage

```python
from transformers import AutoProcessor

processor = AutoProcessor.from_pretrained("samuelstevens/LFM2.5-VL-450M-no-tiling")
assert processor.image_processor.do_image_splitting is False
```

For reproducible runs, pass this repository's commit SHA as `revision`.
Load model weights from the same repository. Runtimes that override the saved
processor configuration need separate verification.

The name retains `LFM` and `-VL-` so LQH's current name-based vision routing can
recognize it. This is not a guarantee of cloud execution or training success.

## Attribution and license

Original model by Liquid AI, Inc. The [LFM Open License v1.0](LICENSE) is retained
unchanged, including its commercial-use conditions. This is a configuration fork
published by samuelstevens, not a new official Liquid AI release.

The original model card is preserved verbatim in [UPSTREAM_README.md](UPSTREAM_README.md).
Its examples and benchmark claims refer to the upstream model with upstream
settings, not newly measured results for this no-tiling fork.

`fork_manifest.json` records the pinned source revision and file SHA-256 hashes.