Llama 2 · Meta · Model History

Claims and evidence

Released 18 July 2023 Primary source[1]page date: July 18, 2023; Takeaways [2]v1 submission date

Status AvailableWeights are downloadable from Meta's meta-llama organisation on Hugging Face after accepting the Llama 2 Community License (gated, seen 2026-10-01). Meta's former download page (llama.com/llama-downloads) now redirects to dev.meta.ai and shows no public content. The meta-llama/llama GitHub repository is marked deprecated; that concerns the code repository, not the model. Primary source[3] [4]

Successor of LLaMA Primary source[1]paragraphs on the demand for Llama 1 and the next version [2]Section 1 Introduction; Table 1

Change · Licensing Free for research and commercial use under the Llama 2 Community License, where Llama 1 was offered under a noncommercial research licence.Compared with LLaMA Primary source[1] [4]
Change · Context length Context length doubled from 2k to 4k tokens.Compared with LLaMA Primary source[2]Table 1
Change · Other Released with dialogue-tuned versions, Llama 2-Chat, trained with supervised fine-tuning and reinforcement learning from human feedback.Compared with LLaMA Primary source[2]
Change · Training data Trained on a new mix of publicly available data; the pretraining corpus grew by about 40 percent, to 2 trillion tokens.Compared with LLaMA Primary source[2]Section 1 Introduction; Table 1

Input text Primary source[5]Model Details: Input
Output text Primary source[5]Model Details: Output
Open weights YesWeights for the pretrained and chat models are released under the Llama 2 Community License; download requires accepting it. Primary source[1]paragraph beginning We're now ready to open source [3]
Context window 4k tokens Primary source[5]Model Details table: Context Length [2]Table 1
Access Open-weights download, cloud partnerAt launch: Azure AI model catalog, Amazon Web Services and Hugging Face, among other providers. Primary source[1]paragraph beginning We're now ready to open source

Not derived from another model.

None recorded.

Same dates as Llama 2.

Variant Inline variant in this record.Pretrained and chat-tuned (Llama 2-Chat) versions, each in original and Hugging Face transformers (-hf) format. Primary source[4] [5]Model Details: Variations

Also known as: Llama-2-7b, Llama-2-7b-hf, Llama 2-Chat 7B, Llama-2-7b-chat, Llama-2-7b-chat-hf

Differs in:

Same dates as Llama 2.

Variant Inline variant in this record.Pretrained and chat-tuned (Llama 2-Chat) versions, each in original and Hugging Face transformers (-hf) format. Primary source[6] [5]Model Details: Variations

Also known as: Llama-2-13b, Llama-2-13b-hf, Llama 2-Chat 13B, Llama-2-13b-chat, Llama-2-13b-chat-hf

Differs in:

Same dates as Llama 2.

Variant Inline variant in this record.Pretrained and chat-tuned (Llama 2-Chat) versions, each in original and Hugging Face transformers (-hf) format. The only released size that uses grouped-query attention, per the model card. Primary source[7] [5]Model Details table and caption

Also known as: Llama-2-70b, Llama-2-70b-hf, Llama 2-Chat 70B, Llama-2-70b-chat, Llama-2-70b-chat-hf

Differs in:

AI Radar coverage starts in June 2026; no coverage linked yet.