Llama 3 is the latest large language model developed by Meta AI, the artificial intelligence arm of Facebook's parent company, Meta. It launched in April 2024 and has since reached version 3.1 in July 2024, adding new capabilities, While it is multi-modal, it doesn't yet output images, video, or audio, and instead maintains a strict focus on text-based outputs, though it does offer coding support.
Llama 3 Has Three Models
You can access three versions of Llama 3. The largest and most capable is the Llama 3.1 405B model, which Meta trained on 405 billion parameters (hence "405B"). It launched with the Llama 3.1 update in July 2024 and made Llama the most capable open-source LLM in the world.
The 405B model has a 128,000 token context window, letting it work with far larger datasets, files, and longer prompts than other LLMs while maintaining contextual awareness across extended inputs. This feature is particularly useful when it comes to generating synthetic data to train other large language models, where context is incredibly important.
Along with extensive pre-training on trillions of tokens of data, 405B received extensive fine-tuning and millions of tokens of feedback from human interactions to improve its accuracy and increase its safeguards against malicious use. The 405B model is almost exclusively for high-level research and running on data center hardware, as its size makes it more costly to run.
Llama 3.1 70B is a more compact model trained on a more modest 70 billion parameters. It's faster and leaner than the 405B model, requiring far less processing power to run. Its smaller data set makes it cheaper and faster and allows it to work on some commercial hardware. It retains the large context window of the 405B model but lacks its depth or breadth of abilities. The 70B model can take on medium AI tasks, such as running chatbots for customer support.
Llama 3.1 8B is the leanest of the models, trained on "just" eight billion parameters. It still maintains the 128,000 token context window, but its more limited training data means it's not as accurate, nor is it as capable of performing complicated analytics as the other models. Its lightweight design makes it capable of running locally on consumer hardware, though, making it a great tool for local text generation and for developers looking to build lightweight commercial applications with AI functionality.
Llama 3's Abilities
Llama 3 has proven to be an incredibly capable collection of large language models that can rival the best of the competition, including the iconic GPT-4 and GPT-4o models.
In benchmarks with other high-end LLMs, Llama 3.1 did very well. Its 405B model is as good or better than most other top models, and its Llama 70B and 8B models are competitive with the leaner, faster models out there.
:max_bytes(150000):strip_icc()/llamacompet-78db02676caa4732a9be3b1dce66b36e.jpg)
Meta
Llama 3 is excellent at coding. It can both help you write your own code or can custom write entire programs for you if you keep it simple. It supports over 30 languages, although it's still best when working with English.
Like most LLM AIs, Llama 3 is fantastic at generating text for writing purposes, too. That goes for business scenarios, fiction, and prose, or social media posts. It can consider a lot of context, too, so you can really fine-tune your prompts.
Llama 3 has more advanced safeguards than previous generations of this LLM. It includes Llama Guard 3, a multi-lingual safety model that works with Prompt Guard, which helps prevent code injection and generation using prompts. These applications are also open-source.
A major advantage of Llama 3 for consumers, developers, and the growth of use of the LLM, is that it's open-source. The code is available to both developers and end users to create something new and better using it as a foundation. This access accelerates the development of Llama 3-powered AIs and chatbots because it's available where other commercially available LLMs can't.
You can even try out the high-end Llama 3.1 405B model yourself by asking Meta.ai or on Whatsapp in the US a challenging math or coding question.
Llama 3's Limitations
Although Llama 3 is incredibly capable, it isn't the best LLM at everything and lacks a significant number of features that you might find elsewhere.
Although it is a multi-modal language model, Llama 3 doesn't have full multi-modal support at the time of writing. Image generation is not widely available, and it can't generate video or audio.
Its capabilities in languages other than English are still limited, too. It has hallucinations and continues to make mistakes confidently, as other AIs of its type have done in the past.
Llama 3.1 is also expensive to run, especially its most capable 405B model, so access to this top model is limited and may remain that way for some time. The other models are impressive, but they don't deliver the same kind of capabilities.
Llama 3 was also incredibly expensive to train. It required the use of thousands of Nvidia H100 GPUs for an extended period of time. So much, in fact, that even a company of the size and available resources as Meta, had to make decisions about when–and for how long–to allow models to use the hardware for training, since they may have needed it elsewhere.
How to Use Llama 3
The easiest way to use Llama 3 is to interact with the Meta.ai chatbot in Facebook Messenger or Whatsapp if you're in the US. That AI uses Llama 3, and depending on the complexity of your task, will let you use Llama 3.1 405B, or the less demanding 70B and 8B models.
You can also use Llama 3 on platforms like HuggingFace that give access to open-source large language models. You'll need to sign up for an account and select Llama 3 from the available models, but then you can use the API to create your own applications or use HuggingFace's libraries to interact with the model directly.
Alternatively, you can download and install GPT4All. It'll let you interact with a range of large language models, including Llama 3 locally. Note, however, that performance when using high-end models is significantly reduced compared to cloud-based LLMs.
No tags.
