Llama-3.2-3B-Instruct

Llama-3.2-3B-Instruct is Meta's lightweight, text-only instruction-tuned model in the Llama 3.2 family, released in September 2024. With 3.21 billion parameters and a 128K-token context window, it targets multilingual dialogue, agentic retrieval, summarization, and on-device assistant use cases where larger models are impractical. It officially supports English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai, and is released under the Llama 3.2 Community License for commercial and research use.

Architecturally, the model uses an optimized autoregressive Transformer with Grouped-Query Attention (GQA) and shared embeddings. Pretraining drew on up to 9 trillion tokens of public data (knowledge cutoff December 2023), with logits distilled from Llama 3.1 8B and 70B teachers to recover performance after pruning. Post-training followed the Llama 3.1 recipe—supervised fine-tuning, rejection sampling, and direct preference optimization—plus RLHF alignment for helpfulness and safety.

Despite its small footprint, the instruct variant scores strongly on common benchmarks for its class: 63.4 on MMLU, 77.7 on GSM8K, 48.0 on MATH, and 77.4 on IFEval. Meta also ships quantized variants (SpinQuant and QLoRA) that cut model size and memory by roughly half while preserving most capability, making this one of the most capable open-weight models for edge and mobile deployment.