News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
SmolLM2

SmolLM2

Compact language models for on-device use

SmolLM2 is a family of small language models released in February 2025 by Hugging Face. It comes in three sizes and focuses on high performance through data-centric training. The largest variant uses 1.7 billion parameters and runs efficiently on local hardware.

Share on WhatsApp

  • Available in 135M, 360M, and 1.7B parameter sizes
  • Trained on 11 trillion tokens of diverse data
  • Outperforms Llama-1B and Qwen2.5-1.5B in benchmarks
  • Supports function calling and instruction following
  • All training datasets released openly

What it is

SmolLM2 is a set of compact language models designed to run on devices with limited resources. Hugging Face released the family in early 2025. It includes versions with 135 million, 360 million, and 1.7 billion parameters. The project aims to prove that small models can achieve strong results if trained on high-quality data. The largest model handles tasks like reasoning, coding, and following instructions. It serves as a lightweight alternative to larger systems that require significant computing power.

How it came about

The team at Hugging Face developed SmolLM2 to improve upon the original SmolLM series. They published their findings in a paper titled "SmolLM2: When Smol Goes Big" in February 2025. The developers focused on a data-centric approach rather than just increasing model size. They created new datasets to address gaps in existing training material. This process involved manual refinement and ablation studies to optimize dataset mixing. The release includes both the model weights and the curated datasets for public use.

Inside the model

The models use a decoder-only transformer architecture. Training involved processing approximately 11 trillion tokens of text. The data mix included web text from FineWeb-Edu and DCLM, plus code from The Stack. The team added specialized datasets for mathematics and instruction following. They introduced new collections like FineMath, Stack-Edu, and SmolTalk. Post-training used supervised fine-tuning and Direct Preference Optimization. This alignment step utilized public datasets and UltraFeedback to improve response quality.

What it can do

The 1.7 billion parameter version shows improvements over its predecessor in knowledge and math. It supports text rewriting, summarization, and function calling. Benchmarks indicate it outperforms similar small models like Llama-1B and Qwen2.5-1.5B in several tests. For example, it scores higher on HellaSwag and ARC evaluations. The instruct model achieves a 27 percent score on the Berkeley Function Calling Leaderboard. It handles multi-turn conversations and specific tool-use tasks effectively.

How to use it

Hugging Face provides open access to the model weights. Users can download the models for local inference on CPUs or GPUs. The company offers code examples for using the models with Transformers and TRL libraries. Developers can also run the models in browsers using Transformers.js. All training datasets created during the project are released alongside the models to support further research.

Why it matters

SmolLM2 demonstrates that small language models can compete with larger counterparts through better data. By releasing the training datasets, the project encourages reproducible research in efficient model development. It provides a baseline for future work on resource-constrained artificial intelligence applications.

Questions people ask about SmolLM2

What sizes does SmolLM2 come in?

The family includes three models with 135 million, 360 million, and 1.7 billion parameters. This range allows users to choose a size that fits their specific hardware constraints.

How much data was used for training?

The models were trained on approximately 11 trillion tokens. This data included web text, code, mathematics, and specialized instruction-following datasets curated by the team.

Can SmolLM2 run on my computer?

Yes, the models are designed for on-device use. They can run on CPUs or GPUs using standard libraries like Transformers, making them suitable for local deployment without cloud services.

Is the code and data open source?

Hugging Face released the model weights and all curated datasets openly. This includes FineMath, Stack-Edu, and SmolTalk, allowing others to reproduce the training process.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on SmolLM2 yet, so the facts on this page come from the sources listed above: huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct, arxiv.org/abs/2502.02737. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.