Llama-2-7b-hf

Llama-2-7b-hf is the 7-billion-parameter base pretrained model from Meta's Llama 2 family, released in July 2023 as part of the first major open-access LLM suite with commercial licensing. It is a decoder-only Transformer trained on 2 trillion tokens of publicly available online data, doubling the context length of Llama 1 to 4,096 tokens and delivering substantial gains across reasoning, knowledge, coding, and safety benchmarks compared to its predecessor.

On standard academic evaluations reported in the official model card and paper, Llama 2 7B achieves 45.3% on MMLU, 32.6% on BBH, 29.3% on AGIEval, 14.6% on GSM8K, and 12.8% pass@1 on HumanEval. It outperforms Llama 1 7B and contemporary open models such as MPT-7B and Falcon-7B on most grouped benchmarks, while remaining lightweight enough for single-GPU deployment.

The model is intended as a foundation for downstream fine-tuning and research rather than direct conversational use. Meta also released instruction-tuned Llama-2-7b-chat-hf variants optimized for dialogue. Weights are distributed via Hugging Face under the Llama 2 Community License, requiring acceptance of Meta's license terms before download.