README.md

metadata

title: Demo of Encoder-Free VLM Trained for $100
emoji: 🖼️
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: apache-2.0
short_description: Ask questions about images and get answers instantly

Encoder-Free VLM Demo

An encoder-free vision-language model: there is no pretrained vision encoder. A learned patch embedder splits a 512×512 image into 32×32 patches and projects them directly into the hidden space of Qwen3-1.7B. The projected patch embeddings are spliced into <|image|> placeholder positions and the decoder generates from inputs_embeds.

This Space serves the best checkpoint trained on the DenseFusion-1M + ShareGPT4V data mix.

Upload a single image, ask a question, and the model streams a response. The demo is single-image and stateless — each turn is a fresh conversation.

Inspired by the SmolVLM-256M-Demo Space.