SmolVLM2
Tiny vision-language models (256M–2.2B) for image and video on edge hardware.
About
Hugging Face's TB Research team built SmolVLM2 in three sizes — 256M, 500M, and 2.2B — for image and video understanding on consumer hardware: the 2.2B can summarize video on an RTX 3060 using 5.2GB of VRAM. It pairs a SigLIP image encoder with a SmolLM2 text decoder, trained on 3.3M samples across ten datasets, and ships Apache 2.0.
Independent testing found it solid on OCR and object counting but weak on zero-shot object detection — know its limits.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Hugging Face TB Research team; announced Feb 20, 2025; 256M/500M/2.2B sizes; SigLIP encoder + SmolLM2 decoder; 3.3M training samples across ten datasets; Independent Roboflow testing: strong on OCR/counting, weak on zero-shot object detection. License: Apache 2.0 — full commercial use permitted..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Hugging Face TB Research team; announced Feb 20, 2025; 256M/500M/2.2B sizes
- · SigLIP encoder + SmolLM2 decoder; 3.3M training samples across ten datasets
- · Independent Roboflow testing: strong on OCR/counting, weak on zero-shot object detection