Home › Language Models › SmolLM

SmolLM - AI Language Models Tool

Overview

SmolLM is an open-source family of compact, efficiency-focused transformer models from Hugging Face designed for on-device and edge deployment across text and vision tasks. The project includes small-scale language models (commonly cited sizes: ~135M, ~360M and ~1.7B) and a later SmolLM3 3B variant that Hugging Face published with full training recipes, weights, and evaluation results. SmolLM3 emphasizes long-context reasoning, multilingual support, and instruction-tuned “dual-mode” reasoning while remaining compact enough for many production and research deployments. ([github.com](https://github.com/huggingface/smollm)) The family also includes SmolVLM, a vision–language line built on the same efficiency principles for tasks such as visual QA, image description and multi-image dialogues. The SmolLM releases are fully open (Apache-2.0) and come with detailed training data mixtures, ablation results, and examples for local inference and several runtime backends (transformers, vLLM, GGUF/llama.cpp, ONNX). SmolLM3 in particular was trained on an 11+ trillion token curriculum and supports extreme context lengths (trained at 64k and extrapolated to 128k tokens using YaRN/NoPE strategies), making it a practical option when long-document understanding and on-device efficiency are both priorities. ([huggingface.co](https://huggingface.co/blog/smollm3?utm_source=openai))

GitHub Statistics

  • Stars: 3,638
  • Forks: 276
  • Contributors: 21
  • License: Apache-2.0
  • Primary Language: Python
  • Last Updated: 2026-01-13T15:59:49Z

The smollm repository is active and well-attended: the project is published under Apache-2.0 and the repo shows ~3.6k stars and ~276 forks on GitHub, with dozens of issues and ongoing contributions. The maintainers provide comprehensive README material, links to model checkpoints and model-hub pages, and a set of local-inference utilities (text/vision/tools folders) that lower the barrier to reproduce and run models locally. Activity indicators (issues, PRs, and recent commits) plus published model cards and Hugging Face blog coverage point to strong engineering investment and community interest. (Repository metadata and README). ([github.com](https://github.com/huggingface/smollm))

Installation

Install via pip:

git clone https://github.com/huggingface/smollm.git
pip install -U transformers
pip install vllm  # optional: for vLLM runtime integration
python -c "from transformers import AutoModelForCausalLM, AutoTokenizer; tokenizer=AutoTokenizer.from_pretrained('HuggingFaceTB/SmolLM3-3B'); model=AutoModelForCausalLM.from_pretrained('HuggingFaceTB/SmolLM3-3B')"

Key Features

  • Compact model lineup: common sizes include ~135M, ~360M and ~1.7B parameters (plus SmolLM3 at 3B).
  • SmolLM3 — 3B model with open weights and full training recipe (released July 2025).
  • Long-context support: trained at 64k tokens and extrapolated to 128k using YaRN/NoPE.
  • Grouped-Query Attention (GQA) and NoPE used to reduce KV cache and improve long-context efficiency.
  • Instruction-tuned dual-mode reasoning (think / no_think) for higher-effort and fast modes.
  • SmolVLM: a multimodal (vision+language) 2B-class model for image QA and multi-image conversations.

Community

SmolLM has broad community visibility: the GitHub repo (Apache-2.0) has thousands of stars and hundreds of forks, active issues and pull requests, and published model cards on Hugging Face’s Model Hub. Hugging Face published a detailed SmolLM3 blog and docs, which spurred coverage and discussion on forums and social platforms (Reddit, LinkedIn, community blog posts) and integrations by community members (examples using Ollama, llama.cpp, GGUF, ONNX and vLLM). The project is used as a base for local/edge inference experiments and small-model research; community feedback highlights strong benchmark performance for its size and many practical examples of on-device deployment. ([github.com](https://github.com/huggingface/smollm))

Last Refreshed: 2026-03-03

Key Information

  • Category: Language Models
  • Type: AI Language Models Tool
Visit Official Website

Related Tools

Qwen2.5-7B

Qwen2.5-7B is a large language model designed for text generation, featuring improvements in coding, mathematics, instruction following, long text generation, and multilingual support. It supports context lengths up to 128K tokens and is intended for sophisticated NLP tasks.
DeepSeek-V3

Open‑weight flagship LLM (base and chat) with strong reasoning and tool‑use; runs with SGLang/LMDeploy/vLLM and permissive model license.
Llama 3

Llama 3 is an open access large language model (LLM) released by Meta, available in various configurations (8B and 70B parameters) with capabilities for fine-tuning and integrations into platforms like Hugging Face, Google Cloud, and Amazon SageMaker.
UNfilteredAI-1B

A large-scale text generation model designed for creative and unconstrained content production without traditional filtering.
Shuttle-3

Shuttle-3 is a state-of-the-art language model designed for high-quality text generation, particularly suited for complex chat, multilingual communication, and reasoning tasks. It is fine-tuned from the Qwen-2.5-72b-Instruct model and designed to emulate high-quality prose similar to Claude 3 models.
WizardLM

WizardLM is a state-of-the-art large language model designed for complex chat, multilingual tasks, reasoning, and agent functionalities. It features an AI-powered pipeline (Auto Evol-Instruct) that optimizes instruction datasets for improved performance across various domains and leverages Arena Learning for an expanded learning pool of challenging instruction data.