SmolVLM

Claims and evidence

• Released 26 November 2024 ● Primary source[1]page date

• Status Available ● Primary source[2]model repository (weights hosted on the Hub)

• Derived from (other) SmolLM2SmolLM2 serves as the language backbone of the vision-language model (the 1.7B Instruct checkpoint after a context extension to 16k tokens); the 256M and 500M variants use SmolLM2-135M-Instruct and SmolLM2-360M-Instruct. ● Primary source[1]Architecture: SmolLM2 1.7B as the language backbone [2]metadata base_model: HuggingFaceTB/SmolLM2-1.7B-Instruct

• Change · Architecture Uses SmolLM2 1.7B instead of Llama 3.1 8B as the language backbone, within the Idefics3 architecture.Compared with Idefics3 ● Primary source[1]Architecture
• Change · Efficiency Compresses visual information 9x with pixel shuffle, compared with 4x in Idefics3.Compared with Idefics3 ● Primary source[1]Architecture
• Change · Size About 2B parameters in total, with a 1.7B language backbone in place of an 8B one.Compared with Idefics3 ● Primary source[1]

• Input text, image ● Primary source[2]introduction; Model Summary: Model type
• Output text ● Primary source[2]introduction
• Open weights Yes ● Primary source[1]TLDR: checkpoints released under Apache 2.0 [2]Model Summary: License
• Context window 16k tokens tokensContext of the SmolLM2 backbone after extension, for the 2B model; the 256M and 500M variants use 8k. ● Primary source[1]Training Details: Context extension
• Parameters 2BThe November 2024 model; later Hugging Face pages give 2.2B. ● Primary source[1]TLDR
• Access Open-weights download, On device ● Primary source[2]introduction: suitable for on-device applications [1]What is SmolVLM?

Lineage

Variants

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Hugging Face on AI Radar →