Llama 3 70B Instruct

Meta Llama 3 70B Instruct is the instruction-tuned 70B model in Meta's first Llama 3 release. According to the Hugging Face model card, Meta released it on April 18, 2024. It is a text model tuned for assistant-style chat, and the card now notes that Llama-3.3-70B-Instruct is available as a newer option. Meta's Llama 3 announcement introduced the 8B and 70B sizes, in pretrained and instruction-tuned versions, with larger models of over 400B parameters still training at the time.

Model Specs

Key Features

• Assistant-style chat: the card says it is intended for commercial and research use in English, and use in other languages is out of scope unless developers fine-tune it.
• Vendor benchmarks: according to Meta's card, the instruction-tuned 70B scores 82.0 on MMLU (5-shot), 39.5 on GPQA (0-shot), 81.7 on HumanEval (0-shot), 93.0 on GSM-8K (8-shot, CoT) and 50.4 on MATH (4-shot, CoT). These are Meta's own 2024 numbers.
• Broad availability at launch: Meta listed AWS, Google Cloud, Microsoft Azure, Hugging Face and Databricks among the platforms offering it.
• Open weights: the weights can be downloaded under the community license, which makes self-hosting possible.

Use Cases

• Self-hosted English chat and assistant prototypes.
• A baseline for fine-tuning experiments on a well-known open 70B architecture.
• Reproducing 2024 benchmark comparisons.

A limitation: the 8k context window is short by current standards, and the knowledge cutoff of December 2023 means it lacks recent information. Meta's card also points to Llama 3.3 70B Instruct as the newer option.

Getting Started

1. Accept the Llama 3 Community License on the model card to get access.
2. Plan for multi-GPU inference at 70B scale, or use a hosted provider that serves it.
3. Format prompts with the model's chat template, and test on your own evals before shipping.
4. Compare with newer Llama releases before starting a new project.

FAQ

What is Llama 3 70B Instruct?

Meta's instruction-tuned 70B model from the April 18, 2024 Llama 3 release, built for assistant-style chat.

What is its context length?

8k tokens, per the model card.

Is it free for commercial use?

Meta's Llama 3 Community License allows commercial use, except that companies above 700 million monthly active users must ask Meta for a license.

What replaced it?

The model card points to Llama-3.3-70B-Instruct. In this catalog, see Llama 3.1 70B Instruct.

Alternatives

• Llama 3.1 70B Instruct: a later 70B Llama in this catalog.
• Llama 3 8B Instruct: the 8B sibling for smaller hardware.
• Llama 3.1 405B Instruct: Meta's largest Llama 3.1 model.
• Mistral Small: a smaller open-weight alternative from Mistral.

Conclusion

Llama 3 70B Instruct was a milestone open-weight model in 2024, and it is still available for self-hosting under Meta's license. For new work, compare it with the newer Llama releases the model card points to. Start from the official Llama site and see more in the models category.