Llama 2
A series of open-weight large language models released by Meta, available in multiple parameter sizes including 7B, 13B, and 70B. It supports local deployment and is among the first open-source LLMs to achieve broad commercial adoption.
Core Facts
When Meta previously released LLaMA 2 and LLaMA 3, it used a custom community license that set additional authorization thresholds for companies with more than 700 million monthly active users.
Meta's Llama series license explicitly prohibits use by commercial products with more than 700 million monthly active users and requires derivative models to retain brand identifiers.
CodeLlama was trained by Meta through continued pre-training and instruction fine-tuning of Llama 2 in the code domain, offered in multiple parameter scales including 7B, 13B, and 34B.
Timeline (last 90 days)
Latxa is a family of Basque large language models released by the University of the Basque Country (UPV/EHU) and other institutions, built by continued pre-training and instruction fine-tuning of open-weight models such as Llama 2 on Basque corpora.
The flagship version of Llama 3 was pretrained on 15.6 trillion tokens, with roughly 50 times the compute of the largest Llama 2.
Llama 2 used the AdamW optimizer (beta1 = 0.9, beta2 = 0.95, weight decay 0.1, gradient clipping 1.0), and was pretrained on 2 trillion tokens.
GQA compresses the KV Cache by about 8x relative to MHA and has become the standard choice for mainstream open-source models such as Llama 2/3, Qwen, and Mistral.
The releases of Llama 2 and Llama 3 broke the monopoly of closed-source models over high-performance AI capabilities.
Llama 3 relaxed the license restrictions to some extent compared to Llama 2, but it is still not an OSI-certified standard open-source license.
Meta has made the weights of Llama 2/3 publicly available, but with specific commercial use restrictions.
The LLaMA leak prompted Meta to adjust its licensing strategy in the subsequent releases of Llama 2 and Llama 3.
Llama 2 is one of the first commercially successful open-weight models, released by Meta.
Llama 2 is available in multiple versions with 7 billion, 13 billion, and 70 billion parameters.
9 more timeline events
All Facts (20)
When Meta previously released LLaMA 2 and LLaMA 3, it used a custom community license that set additional authorization thresholds for companies with more than 700 million monthly active users.
75%VerifiedMeta's Llama series license explicitly prohibits use by commercial products with more than 700 million monthly active users and requires derivative models to retain brand identifiers.
65%VerifiedCodeLlama was trained by Meta through continued pre-training and instruction fine-tuning of Llama 2 in the code domain, offered in multiple parameter scales including 7B, 13B, and 34B.
65%UnverifiedMistral的7B模型在发布时超越了Llama 2 13B的表现
85%UnverifiedMeta's Llama 3 made extensive use of synthetic data generated by Llama 2 during training to enhance specific capabilities.
70%UnverifiedLlama 2's chat template uses [INST] and [/INST] markers to delimit user instructions, and <<SYS>> and <</SYS>> markers to delimit system prompts.
60%UnverifiedThe latest version of Meta's Llama series is Llama 4.
60%UnverifiedLatxa is a family of Basque large language models released by the University of the Basque Country (UPV/EHU) and other institutions, built by continued pre-training and instruction fine-tuning of open-weight models such as Llama 2 on Basque corpora.
50%UnverifiedThe flagship version of Llama 3 was pretrained on 15.6 trillion tokens, with roughly 50 times the compute of the largest Llama 2.
50%UnverifiedLlama 2 used the AdamW optimizer (beta1 = 0.9, beta2 = 0.95, weight decay 0.1, gradient clipping 1.0), and was pretrained on 2 trillion tokens.
50%UnverifiedGQA compresses the KV Cache by about 8x relative to MHA and has become the standard choice for mainstream open-source models such as Llama 2/3, Qwen, and Mistral.
50%UnverifiedThe releases of Llama 2 and Llama 3 broke the monopoly of closed-source models over high-performance AI capabilities.
50%UnverifiedLlama 3 relaxed the license restrictions to some extent compared to Llama 2, but it is still not an OSI-certified standard open-source license.
50%UnverifiedMeta has made the weights of Llama 2/3 publicly available, but with specific commercial use restrictions.
50%UnverifiedThe LLaMA leak prompted Meta to adjust its licensing strategy in the subsequent releases of Llama 2 and Llama 3.
50%UnverifiedLlama 2 is available in multiple versions with 7 billion, 13 billion, and 70 billion parameters.
50%UnverifiedLlama 2 is one of the first commercially successful open-weight models, released by Meta.
50%UnverifiedThe Llama series is a family of large language models that Meta has been open-sourcing since 2023, including versions such as Llama 2 and Llama 3, with parameter scales ranging from 7 billion to 70 billion.
50%UnverifiedMeta's Llama 2 (70-billion-parameter version) used 1.8 million GPU hours for training.
50%UnverifiedThe GQA technique was initially proposed by Google Research in 2023 and was subsequently widely adopted by mainstream models such as Llama 2 and Mistral.
50%