Key takeaways

• Llama 3.2 3B Instruct is designed for ultra-efficient deployment on edge devices and real-time applications, despite its small size (3B parameters) and Transformer architecture.
• The model demonstrates strong factual recall, achieving 78% accuracy on the easy subset of the AI2 Reasoning Challenge, showing consistent performance and ability to identify correct answers based on simple reasoning.
• The model struggles with deep science, complex math, multi-step problem-solving, and retrieving specialized factual data, with common failures like incorrect application of scientific knowledge and misinterpreting questions, as evidenced by its 0% accuracy on the AIME 2025 Problems Benchmark.
• It achieved suboptimal performance (50.86% accuracy) on the SimpleQA benchmark, indicating significant challenges in accessing or retrieving specific factual data and over-relying on general knowledge.
• Despite its efficiency and low latency, improvements are needed in factual accuracy and providing comprehensive responses, particularly in specialized domains, to address its common failure modes of inaccurate or incomplete information.

Technical specifications

Context length

131k tokens

Max output

—

Architecture

Transformer

Modalities

text->text

Parameters

~3B

Open weights

Yes

Licence

Meta AUP

Regions

USA

Vendor

Meta

Provider model key

meta-llama/llama-3.2-3b-instruct