Llama 3.1

// Overview

Llama 3.1 is Meta's family of open-weight large language models, released on July 23, 2024, in three sizes: 8B, 70B, and 405B parameters. Its headline release, Llama 3.1 405B, was the first openly available model to rival top proprietary systems of its time — Meta benchmarked it as competitive with GPT-4, GPT-4o, and Claude 3.5 Sonnet on general knowledge, math, tool use, and multilingual tasks. All three sizes are dense decoder-only transformers, pretrained on more than 15 trillion tokens, with the 405B trained on a cluster of over 16,000 NVIDIA H100 GPUs.

Compared with the original Llama 3 (8B and 70B, April 2024), Llama 3.1 expands the context window to 128K tokens across all sizes and adds official multilingual support for eight languages: English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai. The models use Grouped-Query Attention for efficient inference, and the instruction-tuned variants were aligned with supervised fine-tuning and RLHF. The knowledge cutoff is December 2023. Notably, Meta stuck with a dense architecture rather than mixture-of-experts to keep training stable at the 405B scale.

Llama 3.1 is distributed under the Llama 3.1 Community License, a custom commercial license that — for the first time in the Llama line — explicitly permits using model outputs to improve other models. The weights are downloadable from Meta and Hugging Face, and the models are hosted by AWS Bedrock, Azure AI, Google Cloud Vertex AI, Together AI, Fireworks AI, and others. Llama 3.1 has since been superseded: Llama 3.3 70B (December 2024) matched the 405B on many tasks at a fraction of the size, and Llama 4 (April 2025) moved the line to a mixture-of-experts design.

// Benchmarks

1. MMLU (5-shot)87.3%
2. MMLU-Pro (CoT)73.3%
3. GPQA (0-shot)50.7%
4. HumanEval (0-shot)89%
5. MATH (CoT)73.8%
6. IFEval88.6%
7. BFCL (tool use)88.5%
8. Multilingual MGSM (CoT)91.6%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Llama 3.1 is open-weight with no single official price. These are the 405B Instruct median figures across hosted API providers; the 8B and 70B variants are substantially cheaper, and self-hosting incurs only compute cost.

Pricing source ↗

// Strengths

• First open-weight model to reach frontier-class performance — the 405B scored 87.3% on MMLU, matching GPT-4-class proprietary systems
• Fully downloadable weights under a permissive commercial license, enabling self-hosting, fine-tuning, and on-prem deployment
• 128K-token context window across all three sizes (8B, 70B, 405B)
• Strong coding (89.0% HumanEval) and math (73.8% MATH) for the 405B Instruct model
• Built-in support for eight languages and strong tool-use / function-calling scores
• Broad ecosystem support — hosted by every major cloud and inference provider

// Best for

• Self-hosted and on-prem deployments where data cannot leave the org's infrastructure
• Fine-tuning a strong open base model on domain-specific data
• Generating synthetic data and distilling smaller models (explicitly permitted by the license)
• Multilingual chat assistants and translation across the eight supported languages
• Long-document summarization and retrieval-augmented generation using the 128K window
• Coding assistants and tool-using agents built on open weights

// How to access

// Llama 3 — every version

The full lineage of the Llama 3 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.