<p>Hugging Face has unveiled SmolLM3, an innovative language model designed for local deployment, featuring 3 billion parameters. This release marks a significant advancement in the domain of agentic applications, boasting impressive reasoning capabilities and extensive multilingual support.</p>
<h3>Key Features and Performance</h3>
<p>SmolLM3 allows users to toggle reasoning capabilities on and off, facilitating a flexible approach to various tasks. It can handle a context length of up to 128K using NoPE and YaRN techniques, making it suitable for processing substantial amounts of information efficiently. The model has demonstrated strong performance in reasoning tasks and tool use, establishing itself as a competitive option alongside larger model architectures.</p>
<h3>Multilingual and Tool Calling Support</h3>
<p>One of the standout features of SmolLM3 is its multilingual capabilities across six languages. It also supports function calling, making it an excellent tool for executing complex tasks. This model has outperformed contemporaries like Llama-3.2-3B and Qwen2.5-3B, while also rivaling larger 4 billion parameter models.</p>
<h3>Training Insights</h3>
<p>Hugging Face has shared a comprehensive training blueprint for SmolLM3, revealing that the total training cost was around $200,000, involving 384 H100 GPUs over 24 days. The three-stage pretraining approach on an impressive 11 trillion tokens, complemented by mid-training enhancements for long-context processing and reasoning improvements, culminates in a post-training phase including supervised fine-tuning and Anchored Preference Optimization.</p>
<h3>Local Deployment</h3>
<p>SmolLM3 is tailored for local execution without the need for proprietary models, offering users autonomy in running AI agents. This capability is further supported by compatibility with transformers v4.53.0, which ensures a seamless integration within local environments.</p>
<p>In conclusion, SmolLM3 emerges as a versatile and powerful model suitable for diverse applications in AI, proving that effective, local language models do not require exorbitantly large architectures to succeed.</p>


Collections

Hugging Face released SmolLM3, a 3 billion parameter language model optimized for local deployment and AI agent applications. The model features toggleable reasoning capabilities, 128K context length support, multilingual functionality across six languages, and function calling abilities. It outperforms similar-sized models like Llama-3.2-3B and Qwen2.5-3B while being trained for $200,000 using 384 H100 GPUs over 24 days on 11 trillion tokens. SmolLM3 enables autonomous AI agent execution without requiring proprietary models.

SmolLM3: A Robust Language Model for Local AI Agents