Llama-2-13b-chat-hf

Llama-2-13b-chat-hf is Meta's 13-billion-parameter instruction-tuned chat model from the Llama 2 family, released in July 2023. It builds on the Llama 2 13B base model, which was pretrained on 2 trillion tokens with a 4,096-token context window—double that of Llama 1—and uses standard decoder-only Transformer architecture with rotary positional embeddings. The chat variant is optimized for multi-turn dialogue through supervised fine-tuning and reinforcement learning from human feedback (RLHF), with additional safety tuning to reduce harmful outputs.

On standard academic benchmarks, the underlying Llama 2 13B base achieves 54.8 on MMLU, 39.4 on BIG-Bench Hard, and 39.1 on AGIEval, with 28.7 on GSM8K and 18.3 pass@1 on HumanEval. The chat-tuned model scores 62.18 on TruthfulQA for truthful and informative generations. Compared to Llama 1 13B, Llama 2 13B shows substantial gains across reasoning, knowledge, and coding tasks, and Llama-2-Chat models were competitive with leading open-source chat models at release.

The model is released under the Llama 2 Community License, permitting commercial use subject to license terms. It requires roughly 26 GB VRAM for full FP16 inference on a single GPU, though 4-bit quantized versions can run on consumer hardware with ~9 GB VRAM. It supports English text input only and integrates natively with the Hugging Face Transformers ecosystem.