--- library_name: transformers license: other license_name: lfm1.0 license_link: LICENSE language: - en - ja - fr - es - de - it - pt - ar - zh - ko pipeline_tag: image-text-to-text tags: - liquid - lfm2 - lfm2-vl - edge ---
Liquid AI
# LFM2‑VL **LFM2-VL-3B** is the newest and most capable model in [Liquid AI](https://www.liquid.ai/)'s multimodal **LFM2-VL** series, designed to process text and images with variable resolutions. Built on the [LFM2](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38) backbone, it extends the architecture for higher-capacity reasoning and stronger visual understanding while retaining efficiency. We are releasing the weights of the new [3B](https://huggingface.co/LiquidAI/LFM2-VL-3B) checkpoint—offering higher performance across benchmarks while remaining optimized for scalable deployment. * **Competitive multimodal performance** among lightweight open models. * **Enhanced visual understanding and reasoning**, particularly on fine-grained perception tasks * **Retains efficient inference** with the same flexible architecture and user-tunable speed-quality tradeoffs * **Processes native resolutions up to 512×512** with intelligent patch-based handling for larger inputs For more details, see the [LFM2-VL-3B post](https://www.liquid.ai/blog/lfm2-vl-3b-a-new-efficient-vision-language-for-the-edge) and the [LFM2 blog post](https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models). ## 📄 Model details Due to their small size, **we recommend fine-tuning LFM2-VL models on narrow use cases** to maximize performance. They were trained for instruction following and lightweight agentic flows. Not intended for safety‑critical decisions. | Property | [**LFM2-VL-450M**](https://huggingface.co/LiquidAI/LFM2-VL-450M) | [**LFM2-VL-1.6B**](https://huggingface.co/LiquidAI/LFM2-VL-1.6B) | [**LFM2-VL-3B**](https://huggingface.co/LiquidAI/LFM2-VL-3B) | |---|---:|---:|---:| | **Parameters (LM only)** | 350M | 1.2B | 2.6B | | **Vision encoder** | SigLIP2 NaFlex base (86M) | SigLIP2 NaFlex shape-optimized (400M) | SigLIP2 NaFlex large (400M) | | **Backbone layers** | hybrid conv+attention | hybrid conv+attention | hybrid conv+attention | | **Context (text)** | 32,768 tokens | 32,768 tokens | 32,768 tokens | | **Image tokens** | dynamic, user-tunable | dynamic, user-tunable | dynamic, user-tunable | | **Vocab size** | 65,536 | 65,536 | 65,536 | | **Precision** | bfloat16 | bfloat16 | bfloat16 | | **License** | LFM Open License v1.0 | LFM Open License v1.0 | LFM Open License v1.0 | **Supported languages:** English **Generation parameters**: We recommend the following parameters: - Text: `temperature=0.1`, `min_p=0.15`, `repetition_penalty=1.05` - Vision: `min_image_tokens=64` `max_image_tokens=256`, `do_image_splitting=True` **Chat template**: LFM2-VL uses a ChatML-like chat template as follows: ``` <|startoftext|><|im_start|>system You are a helpful multimodal assistant by Liquid AI.<|im_end|> <|im_start|>user Describe this image.<|im_end|> <|im_start|>assistant This image shows a Caenorhabditis elegans (C. elegans) nematode.<|im_end|> ``` Images are referenced with a sentinel (``), which is automatically replaced with the image tokens by the processor. You can apply it using the dedicated [`.apply_chat_template()`](https://huggingface.co/docs/transformers/en/chat_templating#applychattemplate) function from Hugging Face transformers. **Architecture** - **Hybrid backbone**: Language model tower (LFM2-2.6B) paired with SigLIP2 NaFlex vision encoders (400M shape-optimized) - **Native resolution processing**: Handles images up to 512×512 pixels without upscaling and preserves non-standard aspect ratios without distortion - **Tiling strategy**: Splits large images into non-overlapping 512×512 patches and includes thumbnail encoding for global context - **Efficient token mapping**: 2-layer MLP connector with pixel unshuffle reduces image tokens (e.g., 256×384 image → 96 tokens, 1000×3000 → 1,020 tokens) - **Inference-time flexibility**: User-tunable maximum image tokens and patch count for speed/quality tradeoff without retraining **Training approach** - Builds on the LFM2 base model with joint mid-training that fuses vision and language capabilities using a gradually adjusted text-to-image ratio - Applies joint SFT with emphasis on image understanding and vision tasks - Leverages large-scale open-source datasets combined with in-house synthetic vision data, selected for balanced task coverage - Follows a progressive training strategy: base model → joint mid-training → supervised fine-tuning ## 🏃 How to run LFM2-VL You can run LFM2-VL with Hugging Face [`transformers`](https://github.com/huggingface/transformers) via installing Transformers from source as follows: ```bash pip install git+https://github.com/huggingface/transformers.git@87be5595081364ef99393feeaa60d71db3652679 pillow ``` Here is an example of how to generate an answer with transformers in Python: ```python from transformers import AutoProcessor, AutoModelForImageTextToText from transformers.image_utils import load_image # Load model and processor model_id = "LiquidAI/LFM2-VL-3B" model = AutoModelForImageTextToText.from_pretrained( model_id, device_map="auto", dtype="bfloat16" ) processor = AutoProcessor.from_pretrained(model_id) # Load image and create conversation url = "https://www.ilankelman.org/stopsigns/australia.jpg" image = load_image(url) conversation = [ { "role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": "What is in this image?"}, ], }, ] # Generate Answer inputs = processor.apply_chat_template( conversation, add_generation_prompt=True, return_tensors="pt", return_dict=True, tokenize=True, ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=64) processor.batch_decode(outputs, skip_special_tokens=True)[0] # This image captures a vibrant street scene in a Chinatown area. The focal point is a large red Chinese archway with gold and black accents, adorned with Chinese characters. Flanking the archway are two white stone lion statues, which are traditional guardians in Chinese culture. ``` You can directly run and test the model with this [Colab notebook](https://colab.research.google.com/drive/11EMJhcVB6OTEuv--OePyGK86k-38WU3q?usp=sharing). ## 🔧 How to fine-tune We recommend fine-tuning LFM2-VL models on your use cases to maximize performance. | Notebook | Description | Link | |-----------|----------------------------------------------------------------------|------| | SFT (TRL) | Supervised Fine-Tuning (SFT) notebook with a LoRA adapter using TRL. | Colab link | ## 📈 Performance | Model | Average | MMStar | RealWorldQA | MM-IFEval | BLINK | MMBench (dev en) | OCRBench | POPE | |-------------------|----------|--------|--------------|------------|--------|------------------|-----------|-------| | InternVL3_5-2B | 66.50 | 57.67 | 60.78 | 47.31 | 50.97 | 78.18 | 834.00 | 87.17 | | Qwen2.5-VL-3B | 65.42 | 56.13 | 65.23 | 38.62 | 48.97 | 80.41 | 824.00 | 86.17 | | InternVL3-2B | 67.44 | 61.10 | 65.10 | 38.49 | 53.10 | 81.10 | 831.00 | 90.10 | | SmolVLM2-2.2B | 56.01 | 46.00 | 57.50 | 19.42 | 42.30 | 69.24 | 725.00 | 85.10 | | LFM2-VL-3B | 69.00 | 57.73 | 71.37 | 51.83 | 51.03 | 79.81 | 822.00 | 89.01 | More benchmark scores are reported in our [LFM2-VL-3B post](https://www.liquid.ai/blog/lfm2-vl-3b-a-new-efficient-vision-language-for-the-edge). We obtained the scores for competitive models using VLMEvalKit. Qwen3-VL-2B is not listed in the results table, as its release occurred the day before. ## 📬 Contact If you are interested in custom solutions with edge deployment, please contact [our sales team](https://www.liquid.ai/contact).