Llama-2-70b-hf

Llama-2-70b-hf is Meta's 70-billion-parameter pretrained base language model from the Llama 2 family, released in Hugging Face Transformers format. It is an auto-regressive text model trained on 2 trillion tokens of publicly available data, intended for research and commercial fine-tuning rather than direct conversational use. Compared with Llama 1 65B, it delivers substantial gains across reasoning, knowledge, math, and coding benchmarks while remaining openly available under the Llama 2 Community License.

Architecturally, Llama 2 70B uses an optimized decoder-only Transformer with Grouped-Query Attention (GQA) to improve inference scalability. The model supports a 4,096-token context window, has 80 layers, and was trained between January and July 2023 on Meta's Research Super Cluster using A100-80GB GPUs. Unlike the Llama-2-Chat variants, this checkpoint is not instruction-tuned or RLHF-aligned and is best suited as a foundation for domain adaptation, supervised fine-tuning, or custom chat pipelines.

On standard academic benchmarks, Llama 2 70B achieves 68.9 on MMLU, 51.2 on BBH, and 54.2 on AGIEval, with strong commonsense and reading comprehension performance relative to contemporaneous open models. It scores 56.8 on GSM8K, 30.5 on HumanEval, and 45.4 on MBPP, reflecting solid but pre-chat-era coding and math capabilities. Developers should pair it with appropriate prompting or fine-tuning for production assistant use cases.