SmolLM3 3B
HuggingFaceTB/SmolLM3-3B
Open Source · chat · open-weights
Open
Alert me on changes
Context
128K
Max output
—
Weights
Open
API $/1M
—
Modalities
text
Released
08 Jul 2025
License: apache-2.0 · HuggingFaceTB/SmolLM3-3B
AI summary
● machine-written
SmolLM3 3B: Open-weights multilingual language model with 128K context
SmolLM3 is a 3-billion-parameter open-source language model trained on 11.2 trillion tokens with support for 6 languages and a 128,000-token context window. The model uses grouped-query attention and depth-over-width architecture optimization to deliver performance comparable to larger models while remaining efficient for edge deployment. It supports both fast responses and extended reasoning modes through /think and /no_think capabilities.
What's new
- 128K token context length with YaRN extrapolation
- Trained on 11.2 trillion tokens with three-stage curriculum mixing web, code, and math data
- Dual-mode intelligence: /think for reasoning and /no_think for fast responses
- Supports 6 languages: English, French, Spanish, German, Italian, Portuguese
- Grouped-query attention (GQA) architecture with rotary embedding (RoPE θ) at 1.5 million
Best for
Edge deployment on mobile and resource-constrained devicesLong-form document analysis and research paper processingMultilingual applications and cross-lingual reasoning tasksTool-calling and intelligent agent applications