Llama 3.1 405B Instruct

Llama 3.1 405B Instruct is the instruction-tuned version of the largest model in Meta's Llama 3.1 family. According to its Hugging Face model card, it was released on July 23, 2024 as a 405B-parameter auto-regressive transformer with Grouped-Query Attention, tuned with supervised fine-tuning and reinforcement learning from human feedback. It takes text in and gives text out, with a 128K-token context.

Status

Available as open weights on Hugging Face under Meta's license; the card is gated. Meta has since released other Llama models, and this catalog lists smaller siblings such as Llama 3.1 70B Instruct and Llama 3.1 8B Instruct. The card does not announce a deprecation.

Key Features

• Scale: 405B parameters (the card's file info lists 406B) and a 128K context length.
• Languages: officially supports 8 languages: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.
• Training data: about 15 trillion tokens of public data for pretraining, with a knowledge cutoff of December 2023. Fine-tuning used public instruction datasets plus over 25 million synthetically generated examples.
• Intended use: commercial and research use, assistant-like chat, and use of its outputs for synthetic data generation and distillation.
• Tool use: the card reports tool-use scores of 92.0 on API-Bank and 88.5 on BFCL.
• Meta-reported benchmarks: MMLU (5-shot) 87.3, MMLU-Pro 73.3, GPQA 50.7, HumanEval 89.0, GSM-8K 96.8 and MATH 73.8. These are self-reported by Meta using its internal evaluation library.

License

The Llama 3.1 Community License requires distributions to prominently display "Built with Llama", requires derived models to include "Llama" at the start of the name, and requires licensees with more than 700 million monthly active users to request a license from Meta. An Acceptable Use Policy applies.

Use Cases

• Synthetic data generation and distillation into smaller models, which the license explicitly allows.
• Self-hosted assistants where open weights and data control matter.
• Research baselines for large open models.

Limitation: a 405B model needs substantial multi-GPU infrastructure, and using it in languages beyond the eight supported needs extra safeguards, per the card.

Pricing

Meta distributes the weights; there is no Meta-hosted price on the model card. Hosted API prices depend on the provider you choose and are not quoted here.

Getting Started

1. Accept Meta's license on the Hugging Face card to get access.
2. Load the weights with Transformers or a serving stack that supports the model, on multi-GPU hardware.
3. For lighter workloads, try Llama 3.1 70B Instruct first.

Frequently Asked Questions

What is Llama 3.1 405B's context length?

128K tokens, according to the model card.

When was it released?

July 23, 2024.

Is Llama 3.1 405B open source?

It is open weights under the Llama 3.1 Community License, which has attribution and naming conditions and a 700 million monthly-active-user threshold. It is not an OSI-style open license.

Which languages does it support?

English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.

Alternatives

• Llama 3.1 70B Instruct: the smaller sibling that is far cheaper to run.
• Llama 3.1 8B Instruct: the smallest Llama 3.1 instruct model.
• DeepSeek V3: another large open-weights chat model.

Conclusion

Llama 3.1 405B Instruct was Meta's largest open chat model at launch, with a 128K context and Meta-reported strong benchmark scores. Read the model card and license before deploying, and consider a smaller Llama 3.1 model if cost matters.