LLaMA 2 is the second generation of a fast and powerfulartificial intelligence (AI)that Meta initially designed for research. Meta officially released LLaMA 2 in 2023, an open source AI model in partnership with Microsoft, so that developers could use the Azure cloud computing platform. It’s also available from other providers.
LLaMA 2 allows AI researchers and commercial developers free use. By comparison, other top-rated large language models like OpenAI’s ChatGPT andGoogle’s AI model Bardare proprietary systems. That makes LLaMA 2 intriguing for third-party developers who want direct access to the model.
What is LLaMA 2?
LLaMA is an acronym for Large Language model Meta AI. Presumably, Meta omitted one ‘M’ from the acronym to make the name more memorable. The first version was introduced in February 2023 in four sizes, containing 7, 13, 33, and 65 billion parameters.
What are large language models?
Large language models (LLMs) are the basis for AI chatbots and much more. Here’s what’s going on behind the scenes
As the number of parameters increases, an AI model becomes more refined, but the degree of sophistication varies among AI models. Meta researchers found LLaMA’s 13B parameter model outperformed OpenAI’s GPT-3, which has 175B parameters. In Meta’s testing, LLaMA-65B performed similarly to the best AI models at that time, Chinchilla-70B and PaLM-540B.
Meta introduced LLaMA 2 in July 2023, the second generation of its AI model. It has a similar design but is trained with a 40% larger data set to create a more advanced AI. Meta surprised the industry by making LLaMA 2 open source code with minimal restrictions on use.
LLaMA 2 comes in three model sizes, from a small but robust 7B model that can run on a laptop and a 13B model suitable for desktop computers to a 70 billion parameter model that requires a powerful computer designed for artificial intelligence research and development.
Qualcomm’s new Snapdragon 8 Gen 3has improved AI processing with several optimized AI models, so the latest Android phones with that chip can run Meta’s AI locally. You can run LLaMA 2 in the cloud for casual use, making this AI available even oninexpensive Android phones.
How good is LLaMA 2?
LLaMA 2 is very good, andMeta saysthe 70B-parameter version can process large text prompts and respond with better performance than ChatGPT (GPT-3.5 Turbo). Meta also released a version of LLaMA 2 optimized for programming,Code LLaMA, that competes well against OpenAI’s GPT-4 when getting assistance with writing code.
While OpenAI and Google have multimodal AI that can accept images as input and create new images, LLaMA 2 only accepts text prompts and generates text replies. For imagery, Meta createdEmu, which works with LLaMA 2 to generate stickers on Instagram.
AI benchmarks are useful in assessing speed and accuracy, and Meta’s LLaMA 2 AI competes well with industry leaders OpenAI and Google. The AI that gives the best results in creative tests will always be subjective.
How can I use LLaMA 2?
TheLLaMA 2 model is available on GitHubas free, open source Python code. Meta requires you torequest accessbefore downloading LLaMA 2’s model weights and tokenizer. These are critical components to run the AI on your computer. You also need a Python interpreter since that’s the language LLaMA 2 is written in.
If any of those requirements sound too difficult, don’t worry. You don’t have to install LLaMA 2 on your computer to use it. Many AI models are free onHugging Face, including LLaMA 2.
You can also try LLaMA 2 on thewebsiteof the same name created by Meta investor and venture capital firm Andreessen Horowitz. It’s free to use in any of the three model sizes, with 7B, 13B, or 70B parameters. Adjusting the number of parameters will give you an idea of how much difference that makes. The 7-billion parameter model is quite good and usable for many tasks.
If you want to use LLaMA 2 for commercial purposes,Microsoft Azure AI Studiocan run LLaMA 2 in the cloud. Alternatively, Amazon’s SageMaker runs on theAmazon Web Servicesplatform.
The great thing about open source AI models is the easy access and diverse options.
LLaMA 2 restrictions
While LLaMA 2 is mostly open source and free for commercial use, Meta has restrictions that could affect developers and commercial users.
You must request a license from Meta if you create an app or a service that becomes popular with over 700 million monthly active users. Your product must become as popular as Snapchat before that becomes a concern.
You cannot use LLaMA 2 output to improve any other large language model. That technique has been used before.Alpacais an open source chatbot created by Stanford University researchers using ChatGPT output. This was strictly done for research, and Alpaca was never deployed.
Best AI apps for your phone
Plenty of apps make it easy to use AI on your Android phone. We havea list of the best Android AI apps, with some tips on how to avoid the clickbait that comes with any concept that creates excitement.
If you live in the US, you can accessLLaMA 2 easily since it’s part of Meta’s WhatsApp and Messenger apps. LLaMA 2 is available globally via a GitHub code repository and online at various AI hosting websites.
