# huggingface/nanoVLM

The simplest, fastest repository for training/finetuning small-sized VLMs.

Repository: https://github.com/huggingface/nanoVLM
Canonical: https://ross.abutalabs.com/products/nanovlm
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2025-10-27T13:16:05+00:00

## Health v2 (maintenance only)
Score: 43/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 49, release rhythm 40, longevity 34
- inputs: {"age_days": 488, "days_push": 310, "days_rel": 455, "gap_med": 15, "n_releases_24m": 2}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 4997, forks 507 (observed 2026-08-28T04:09:04.402783+00:00)

## What it is
nanoVLM is a minimal, readable pure-PyTorch repository (~750 lines) for training and finetuning small Vision-Language Models, inspired by nanoGPT. It combines a SigLIP vision backbone with a SmolLM2 language decoder and includes a simple training loop, primarily serving as an educational codebase.

## Use cases
- train a small vision-language model from scratch
- learn how VLMs work by reading a minimal implementation
- finetune a multimodal model on custom image-text data
- understand how vision and language embeddings are combined
- reproduce a 222M-parameter VLM on a single GPU
- use as a starting codebase for multimodal model experiments

## When to choose
- you want a readable, hackable educational codebase for VLM training
- you have a GPU and want to train a small multimodal model quickly
- you're learning how vision transformers and language decoders connect
- you prefer pure PyTorch over heavy training frameworks

## When to avoid
- you need a production-ready or SOTA multimodal model
- you want polished support scripts, notebooks, and stable APIs (recent breaking changes broke some)
- you need multi-node distributed training out of the box without tweaking
- you want a pretrained VLM for inference rather than training

## Facets
- artifact type: learning-resource
- maturity: active
- function: llm-training, machine-learning, deep-learning
- domain: large-language-models, computer-vision, deep-learning, education
- platform: python
- tags: vision-language-model, pytorch, educational, nanogpt-style, finetuning, multimodal, gpu

## Member repositories
- huggingface/nanoVLM (main) score 43

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:09:04.402783+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:17:36.788170+00:00, confidence not recorded.
  - readme: https://github.com/huggingface/nanoVLM (fetched 2026-08-28T04:09:04.402783+00:00, sha e44d7082f432)
- Data as of 2026-08-30T08:39:29.467469+00:00.