AIAny
AI Model2025
Icon for item

SmolVLM2-256M-Video-Instruct

Answers questions and generates descriptions from video, images, and text. Video inference uses about 1.38 GB of GPU memory, making it suitable for resource-constrained devices and task-specific fine-tuning.

Introduction

Video understanding is often treated as a job for large models, but this checkpoint makes a different trade-off: a 256M-parameter model can handle video questions and descriptions with about 1.38 GB of GPU memory. That makes local experimentation and domain-specific fine-tuning more practical, though its benchmark results reflect the limits of a very small model.

Key Capabilities
  • Accepts video, image, multi-image, and text inputs, then produces text such as captions, answers, or visual narratives. It is an analysis model, not a media generator.
  • Scores 33.7 on Video-MME, 40.6 on MLVU, and 32.7 on MVBench. These results make it a compact baseline for video tasks, not a substitute for larger models when accuracy is paramount.
  • Built on the Idefics3 approach, with SigLIP for visual encoding and SmolLM2 for text generation. Its small footprint supports experiments on constrained hardware and specialized fine-tuning.
  • Released under Apache 2.0, allowing use and adaptation under that license.
Who It Suits—and the Trade-off

Great fit if you need a small, adaptable model for video or image analysis on limited hardware, or want to fine-tune it for a narrow visual task. Look elsewhere if you need high-confidence factual answers, high-stakes decisions, image or video generation, or the strongest benchmark performance: outputs can be inaccurate, and larger variants perform better.

Information

  • Websitehuggingface.co
  • OrganizationsHugging Face
  • AuthorsAndrés Marafioti, Orr Zohar, Miquel Farré, Merve Noyan, Elie Bakouch, Pedro Cuenca, Cyril Zakka, Loubna Ben Allal, Anton Lozhkov, Nouamane Tazi …
  • Published date2025/02/20

Categories

More Items

Hugging Face
AI Model2026

Transcribes speech audio into text across 45 languages, with automatic language detection. Its 0.5B parameters make it a compact option for multilingual transcription using Transformers or vLLM.

Hugging Face
AI Model2026

Combines Korean-English language and image understanding with screenshot-driven browser control in one model. It prioritizes retaining general LLM and VLM capabilities rather than optimizing solely for web tasks.

Hugging Face
AI Model2026

Turns photos into detailed foreground alpha mattes, with particular attention to hair, fur, shadows, and complex edges. Built on a generative image-editing model and distilled to four steps; self-hosted weights are non-commercial only and require substantial GPU memory.