Llama-3.1-405B-Instruct

Llama 3.1 405B Instruct is Meta's flagship open-weight instruction-tuned language model, released in July 2024 as the largest model in the Llama 3.1 family. With 405 billion parameters in a dense decoder-only Transformer architecture using Grouped-Query Attention (GQA), it was the first openly available model to rival leading closed-source frontier models on knowledge, reasoning, math, and coding benchmarks. The model supports a 128K-token context window and eight languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

The instruction-tuned variant is optimized for multilingual dialogue, tool use, and agentic workflows. It was trained with supervised fine-tuning and RLHF on publicly available instruction data plus over 25 million synthetically generated examples. On standard benchmarks it achieves strong results including 88.6% MMLU (CoT), 73.3% MMLU-Pro, 50.7% GPQA, 89.0% HumanEval, 88.6% MBPP, and 96.8% GSM8K, placing it competitively against GPT-4-class models at release.

Llama 3.1 405B Instruct is released under the Llama 3.1 Community License for commercial and research use. It supports native tool-calling formats, structured JSON output, and integration with popular inference frameworks. Deployment requires substantial multi-GPU infrastructure, with FP16 inference needing roughly 810 GB of VRAM for weights alone, making FP8 or INT4 quantization the practical choice for most production deployments.