AI Radar

SmolLM3

Available · Language, Reasoning · Milestone

SmolLM3 is a 3B-parameter open-weight language model from Hugging Face, released under Apache 2.0 as a base model and an instruct model with switchable extended thinking. It natively supports six languages and handles up to 128k tokens of context, and Hugging Face published its architecture details, data mixtures and training configurations. [1] [2] Primary source

Timeline of SmolLM3 →

Claims and evidence

  • Released Primary source[1]page date
  • Status Available Primary source[2]model repository (weights hosted on the Hub)
  • Successor of SmolLM2 Primary source[1]Architecture and training details
  • Change · Reasoning The instruct model adds an extended-thinking mode, on by default and switched with /think and /no_think flags.Compared with SmolLM2 Primary source[1]Enabling and Disabling Extended Thinking Mode
  • Change · Languages Natively supports six languages: English, French, Spanish, German, Italian and Portuguese.Compared with SmolLM2 Primary source[2]Key features
  • Change · Context length Trained on 64k context and handles up to 128k tokens with YaRN, up from 8k for SmolLM2-1.7B.Compared with SmolLM2 Primary source[2]Long context processing [3]section 4.6, Context Length extension
  • Change · Architecture Adds grouped-query attention and NoPE, which removes rotary position embeddings from every fourth layer.Compared with SmolLM2 Primary source[1]Architecture and training details
  • Input text Primary source[2]Model Summary
  • Output text Primary source[2]Model Summary
  • Feature Reasoning modeDual mode: extended thinking on by default, switched with /think and /no_think. Primary source[1]Enabling and Disabling Extended Thinking Mode [2]Key features; Enabling and Disabling Extended Thinking Mode
  • Feature MultilingualSix natively supported languages: English, French, Spanish, German, Italian and Portuguese. Primary source[2]Key features
  • Feature Tool useTool calling through XML tools or Python tools in the chat template. Primary source[2]Agentic Usage [1]Building the Chat Template
  • Open weights Yes Primary source[2]Key features: Fully open model; License
  • Context window 128k tokens tokensTrained on 64k context; up to 128k tokens through YaRN extrapolation. Primary source[2]Key features: Long context; Long context processing [1]Long Context extension
  • Parameters 3B Primary source[2]Model Summary
  • Access Open-weights download Primary source[2]How to use

Lineage

Predecessors

Successors

No known successor.

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

No variants recorded in this record.

Related AI Radar coverage

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

All model releases from Hugging Face on AI Radar →