Home › Vision Models › nanoVLM

nanoVLM

A lightweight, fast repository for training and fine-tuning small vision-language models using pure PyTorch.

Key Information

  • Category: Vision Models
  • Source: Github
  • Last updated: January 09, 2026

Structured Metrics

No structured metrics captured yet.

Links

Visit official website

Canonical source: https://github.com/huggingface/nanoVLM

Related Tools

BLIP-2

BLIP-2 is an advanced visual-language model that allows zero-shot image-to-text generation, enabling tasks such as image captioning and visual question answering using a combination of pretrained vision and language models.
deepfake-detector-model-v1

A deepfake detection image classification model fine-tuned from google/siglip2-base-patch16-512. It leverages the SiglipForImageClassification architecture to classify images as either 'fake' (deepfakes) or 'real', and is intended for applications such as media authentication, content moderation, forensic analysis, and security.
DeepSeek-OCR

DeepSeek-OCR is an open-weight, multilingual vision-language OCR model from DeepSeek that converts images/documents to text (e.g., Markdown) using a context-aware optical compression approach. It runs via Transformers and vLLM, supports FlashAttention, and is provided as BF16 safetensors (~3B params) with example inference code and broad community adoption.
DeepSeek-VL2

A series of advanced vision-language models designed for multimodal understanding, available in multiple sizes to suit varying complexity and performance requirements.
DeepSeek-VL2-small

DeepSeek-VL2-small is a variant of the DeepSeek-VL2 series, advanced mixture-of-experts vision-language models designed for multimodal tasks such as visual question answering, optical character recognition, document/table/chart understanding, and visual grounding.
Florence-2-large

An advanced vision foundation model by Microsoft designed for a wide range of vision and vision-language tasks such as captioning, object detection, OCR, and segmentation. It uses a prompt-based, sequence-to-sequence transformer architecture pretrained on the FLD-5B dataset and supports both zero-shot and finetuned settings.