Llama 3.2 3B

• - 📏 Compact 3B text model with 128K-token context window.
• - ⚡ Built for low-latency, on-device, edge inference.
• - 🔧 Created via structured pruning and distillation from Llama 3.1 8B.
• - 🧠 Pre-trained with logits from Llama 3.1 8B and 70B teachers.
• - 🎯 Supports function-calling/tool use and web search workloads.
• - 🌐 Multilingual: English, German, French, Hindi and more.
• - 🏢 Released by Meta on October 3, 2024.
• - 🔒 Open weights under the Llama 3.2 community license.

Meta

Meta Platforms is an American multinational technology company headquartered in Menlo Park, California. Best known for operating Facebook, Instagram, WhatsApp, and Threads, Meta has become one of the most influential players in open-weight AI research through…

Llama 3.2 3B is one of Meta's smallest, lightweight, text-only Llama models, alongside a 1B sibling, designed to fit onto select edge and mobile devices for tasks like summarization, instruction following, rewriting, and knowledge retrieval running locally. It carries a 128K-token context window and uses grouped-query attention for efficient inference. Meta partnered with hardware vendors including Qualcomm, MediaTek, and Arm to optimize it for modern mobile SoCs.

Compared with earlier full-size Llama models, the 3B was not trained from scratch but produced through structured one-shot pruning from Llama 3.1 8B, followed by knowledge distillation in which logits from the larger Llama 3.1 8B and 70B models served as token-level training targets. Meta describes this as making the 1B and 3B the first highly capable lightweight Llama models that fit on devices efficiently.

Within this catalog's Llama family, the 3B sits at the efficiency end of the spectrum. The much larger Llama 3.3 70B targets higher-quality reasoning and production hosting, while Hermes 3 Llama 3.1 405b is a community fine-tune built atop the 405B base.

The model is openly released under Meta's Llama 3.2 community license, making it suitable for commercial use, embedded systems, and agentic applications where compute and memory are tightly constrained.