> This page is for version Nightly (default).
> For other versions, use one of these documentation indexes:
> - Nightly (default): https://docs.nvidia.com/nemo/automodel/nightly/llms.txt
> - Latest: https://docs.nvidia.com/nemo/automodel/latest/llms.txt
> - 0.7.0 · 26.10: https://docs.nvidia.com/nemo/automodel/v0.7/llms.txt
> - 0.5.0 · 26.06: https://docs.nvidia.com/nemo/automodel/v0.5/llms.txt
> - 0.4.0 · 26.04: https://docs.nvidia.com/nemo/automodel/v0.4/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nemo/automodel/llms.txt.

# LFM3-VL-2B

> Use LFM3-VL-2B with NeMo AutoModel for vision-language fine-tuning, with a documented checkpoint, a runnable recipe, setup guidance, and model reference details.

[LFM3-VL-2B](https://huggingface.co/LiquidAI/LFM3-VL-2B) pairs the LFM3 language model with a vision encoder of gated 2D convolutions and bidirectional attention. It reads images and videos at their native resolution.
The Hugging Face configuration declares the `Lfm3VlForConditionalGeneration` architecture.

## Fine-Tune LFM3-VL-2B

Follow the [installation instructions](/nemo/automodel/get-started/installation), then run the recipe from the repository root:

```bash
automodel examples/vlm_finetune/lfm3/lfm3_vl_2b_medpix.yaml --nproc-per-node 8
```

## Choose a Workflow

| Goal                                   | Start Here                                                                                                                                               |
| -------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Supervised fine-tuning (SFT) on MedPix | Use [lfm3_vl_2b_medpix.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/vlm_finetune/lfm3/lfm3_vl_2b_medpix.yaml). Dataset: MedPix-VQA. |

> **Note**
>
> The checkpoint's processor configuration names image and video processors that exist only as remote code. NeMo AutoModel builds the equivalent Qwen2-VL image and Qwen3-VL video processors from the stored settings, the same mapping Transformers 5.23.0 uses, so `AutoProcessor.from_pretrained` works without `trust_remote_code`. The language model applies the rotary-embedding handling described for [LFM3-1.7B](/nemo/automodel/model-coverage/large-language-models/liquidai/LFM3-1.7B).

## Model Reference

### Model Architecture

| Property                  | Value                                       |
| ------------------------- | ------------------------------------------- |
| Task                      | Image-Text-to-Text                          |
| Architecture              | `Lfm3VlForConditionalGeneration`            |
| Parameters                | 2B                                          |
| Hugging Face Organization | [LiquidAI](https://huggingface.co/LiquidAI) |

### Available Models

| Model      | HF ID                                                               |
| ---------- | ------------------------------------------------------------------- |
| LFM3-VL 2B | [`LiquidAI/LFM3-VL-2B`](https://huggingface.co/LiquidAI/LFM3-VL-2B) |

## Related Resources

* [VLM Fine-Tuning Guide](/nemo/automodel/recipes-e2e-examples/gemma-3-3n)