lusxvr/nanoVLM-222M

nanoVLM is a minimal and lightweight Vision-Language Model (VLM) designed for efficient training and experimentation. Built using pure PyTorch, the entire model architecture and training logic fits within ~750 lines of code. It combines a ViT-based image encoder (SigLIP-B/16-224-85M) with a lightweight causal language model (SmolLM2-135M), resulting in a compact 222M parameter model. The model achieves 35.3% accuracy on MMStar after training for ~6 hours on a single H100 GPU using 1.7M samples from the cauldron dataset, making it a strong baseline for low-resource VLM research.

The model is ideal for researchers and developers interested in exploring VLM training with minimal computational overhead, and serves as a perfect starting point for tinkering with multimodal architectures.

Model Architecture:

• Vision Transformer (SigLIP-B/16)
• Causal Language Model (SmolLM2)
• Modality Projection Layer

Training:

• Trained on ~1.7M samples from the the_cauldron dataset
• 6 hours on a single NVIDIA H100 GPU
• Resulting model size: 222M parameters

Evaluation:

• MMStar Accuracy: 35.3%

Usage:
Usable through the nanoVLM repository: https://github.com/huggingface/nanoVLM
For more details, see: https://github.com/huggingface/nanoVLM?tab=readme-ov-file#hub-integration

from models.vision_language_model import VisionLanguageModel
model = VisionLanguageModel.from_pretrained("lusxvr/nanoVLM-222M")