SyncAI.news, a Varaisys broadcasting
SmolVLM Grows Smaller – Introducing the 256M & 500M Models!
HF

Hugging Face Blog

· 1 min read

AI LabsHugging Face Blog

SmolVLM Grows Smaller – Introducing the 256M & 500M Models!

TLDR

We’re excited to announce two new additions to the SmolVLM family: SmolVLM-256M and SmolVLM-500M. That’s right—256M parameters, making it the smallest Vision Language Model in the world!

We built on everything we learned from SmolVLM 2B while focusing on efficiency, data mixtures, and new design trade-offs. We are excited to introduce a pair of models that preserve strong multimodal performance in a fraction of the footprint.

This release comes with four checkpoints: two base models and two instruction fine-tuned models with sizes 256M and 500M parameters. These models can be loadable directly to transformers, MLX and ONNX, and we have demos for transformers and WebGPU (with ONNX). You can find all the models and the demo for this release here.

Table of Contents

  • Overview
  • Why Go Smaller?
    • Meet the 256M Parameter Giant
    • A Step Up: 500M
  • What Changed Since SmolVLM 2B?
  • Smaller Multimodal Retrieval: ColSmolVLM 256M & 500M
  • Using Smaller SmolVLM
  • Citation information
  • Next Steps

Overview

  • SmolVLM-256M – The world’s smallest VLM!
  • SmolVLM-500M – A half-billion-parameter sibling that offers a significant performance bump while still remaining super lightweight.
  • New Vision Encoder Choices – We compared SigLIP 400M SO (used in SmolVLM 2B and many other large VLMs) against a smaller SigLIP base patch-16/512. Surprisingly, the bigger encoder offered only marginally better results, so we opted for the 93M-parameter SigLIP base patch-16/512 in these new releases.
  • Larger Image Resolution – Our smaller vision encoder processes images at a larger resolution (inspired by Apple’s VLM research and Google’s PaliGemma). This yields sharper image understanding with minimal overhead.
  • Training Optimization – A new tokenization trick significantly boosted real-world benchmarks, even though it made the training loss look worse on paper.

We're now reaching model parity with the SmolLM2 family (135M, 360M, 1.7B), so you have a complete set of smaller LLM + VLM combos to play with.

Original source

This story was published by Hugging Face Blog. SyncAI.news shows a preview; the complete article is on the publisher's site.

Read the full story on huggingface.co

Similar News