Smol Models 🤏

Welcome to Smol Models, a family of efficient and lightweight AI models from Hugging Face. Our mission is to create fully open powerful yet compact models, for text and vision, that can run effectively on-device while maintaining strong performance.

[NEW] SmolLM3 (Language Model)

Our 3B model outperforms Llama 3.2 3B and Qwen2.5 3B while staying competitive with larger 4B alternatives (Qwen3 & Gemma3). Beyond the performance numbers, we're sharing exactly how we built it using public datasets and training frameworks.

Ressources:

• SmolLM3-Base
• SmolLM3
• blog

Summary:

• 3B model trained on 11T tokens, SoTA at the 3B scale and competitive with 4B models
• Fully open model, open weights + full training details including public data mixture and training configs
• Instruct model with dual mode reasoning, supporting think/no_think modes
• Multilingual support for 6 languages: English, French, Spanish, German, Italian, and Portuguese
• Long context up to 128k with NoPE and using YaRN

👁️ SmolVLM (Vision Language Model)

SmolVLM is our compact multimodal model that can:

• Process both images and text and perform tasks like visual QA, image description, and visual storytelling
• Handle multiple images in a single conversation
• Run efficiently on-device

Repository Structure

smollm/
├── text/               # SmolLM3/2/1 related code and resources
├── vision/            # SmolVLM related code and resources
└── tools/             # Shared utilities and inference tools
    ├── smol_tools/    # Lightweight AI-powered tools
    ├── smollm_local_inference/
    └── smolvlm_local_inference/

Getting Started

SmolLM3

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "HuggingFaceTB/SmolLM3-3B"
device = "cuda"  # for GPU usage or "cpu" for CPU usage

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
).to(device)

# prepare the model input
prompt = "Give me a brief explanation of gravity in simple terms."
messages_think = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages_think,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate the output
generated_ids = model.generate(**model_inputs, max_new_tokens=32768)

# Get and decode the output
output_ids = generated_ids[0][len(model_inputs.input_ids[0]) :]
print(tokenizer.decode(output_ids, skip_special_tokens=True))

SmolVLM

from transformers import AutoProcessor, AutoModelForVision2Seq

processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM-Instruct")
model = AutoModelForVision2Seq.from_pretrained("HuggingFaceTB/SmolVLM-Instruct")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "What's in this image?"}
        ]
    }
]

Ecosystem

<div align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/61c141342aac764ce1654e43/RvHjdlRT5gGQt5mJuhXH9.png" width="700"/> </div>

Resources

Documentation

• SmolLM3 Documentation
• SmolLM2 paper
• SmolVLM Documentation
• Local Inference Guide

Pretrained Models

• SmolLM3 Models Collection
• SmolLM2 Models Collection
• SmolVLM Model

Datasets

• SmolLM3 Pretraining dataset
• SmolTalk - Our instruction-tuning dataset
• FineMath - Mathematics pretraining dataset
• FineWeb-Edu - Educational content pretraining dataset