Llama 3.1 70B Instruct

Llama 3.1 70B Instruct is Meta's instruction-tuned 70B model in the Llama 3.1 family, released on July 23, 2024. Its Hugging Face model card lists a 128K context length with grouped-query attention, about 15 trillion tokens of public pretraining data, and a December 2023 knowledge cutoff. It targets assistant-style chat in multiple languages and can also generate synthetic data and support distillation. Weights are available under the Llama 3.1 Community License, which includes an Acceptable Use Policy.

Key features

• 128K context and a 70B dense model, per the card.
• Languages: officially English, German, French, Italian, Portuguese, Hindi, Spanish and Thai; trained on a broader set, and fine-tuning for others is permitted under the license.
• Tool use: supports multiple tool-use formats through chat templates; Meta reports API-Bank 90.0 and BFCL 84.8.
• Meta's benchmarks for the instruct model: MMLU 83.6, MMLU-Pro 66.4, IFEval 87.5, HumanEval 80.5, GSM-8K 95.1, MATH 68.0 and MGSM 86.9. These are vendor figures.
• Fine-tuning data: public instruction datasets plus over 25M synthetic examples.

Use cases

• Multilingual assistants and chat applications.
• Synthetic data generation and distillation into smaller models, which the card lists as supported uses.
• Tool-calling agents on self-hosted infrastructure.

Pricing and availability

The weights are downloadable from Hugging Face after accepting the Llama 3.1 Community License. Meta's card does not list API prices; hosted costs depend on the provider you choose, and self-hosting needs enough GPU memory for 70B parameters (BF16 weights). We checked the card on 2026-10-29 and found no deprecation notice. Uses that violate the license or the Acceptable Use Policy are out of scope.

Getting started

1. Request access on the Hugging Face repo and accept the license.
2. Load it with Transformers or a serving stack such as vLLM, and apply the model's chat template.
3. For tool calling, pass tool definitions through the chat template as described on the card.
4. Test it against newer open models on your own tasks before committing.

FAQ

What is the context length of Llama 3.1 70B Instruct?

128K tokens, according to Meta's model card.

Which languages does it officially support?

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

What is its knowledge cutoff?

December 2023, per the card.

Can I use it commercially?

The Llama 3.1 Community License allows commercial and research use, subject to its terms and the Acceptable Use Policy. Read the license before you deploy.

What about smaller or larger Llama 3.1 models?

See Llama 3.1 8B Instruct and Llama 3.1 405B Instruct.

Alternatives

• Llama 3.1 8B Instruct: the small sibling for modest hardware.
• Llama 3.1 405B Instruct: the largest model in the family.
• NVIDIA Llama 3.1 Nemotron 70B Instruct: NVIDIA's Llama 3.1 Nemotron 70B variant.

Conclusion

Llama 3.1 70B Instruct is a well-documented open-weights chat model with a 128K context and broad tooling support. Review the model card and license, compare it with newer options, and browse the models category.