Meta Introduces New Llama 3 AI Models with Advanced Capabilities

Maryan Duritan
Maryan Duritan
IT Writer
Last updated: May 9, 2024
Why Trust Us
Our editorial policy emphasizes accuracy, relevance, and impartiality, with content crafted by experts and rigorously reviewed by seasoned editors for top-notch reporting and publishing standards.
Disclosure
Purchases via our affiliate links may earn us a commission at no extra cost to you, and by using this site, you agree to our terms and privacy policy.

Meta has introduced the latest additions to its Llama series of AI models, launching two new versions under the Llama 3 umbrella. The goal of this release was to provide developers with more advanced AI development tools. The two models, Llama 3 8B and Llama 3 70B, boast 8 billion and 70 billion parameters respectively. In general, an AI model with more parameters should give better and more accurate results because it can learn from a larger amount of data.

Meta has released these Llama 3 models as open source to support and engage with a broader developer community. This means that any developer can use them for free, which encourages new ideas and could lead to AI-powered apps that are smarter. 

Meta has ambitious plans for the Llama 3 AI Models

The company wants to make the best open-source AI model that can compete with the best private models on the market right now. By making these models open-source, Meta is able to collaborate with developers around the world to continuously improve the models based on collective insights. Llama 3 is currently focused on text processing, but Meta plans to make it multilingual and multimodal as well. They aim to enhance its capabilities to include reasoning and coding, broadening its usefulness across different applications.

Llama 3 models: New performance benchmarks

The latest AI models from Meta, the Llama 3 8B and Llama 3 70B, are a big step up from the Llama 2 line. These new models, which have 8 billion and 70 billion parameters respectively, are noted for their superior performance. Parameters in an AI model determine its ability to analyze and generate text with higher counts generally indicating more capability.

Meta backs up their claims of being better with scores from AI tests such as MMLU (which tests knowledge), ARC (which tests skill acquisition), and DROP (which tests reasoning over text). While the effectiveness of these benchmarks can be debated they are commonly used to assess AI models. 

In comparative tests, Llama 3 8B has outshined comparable models with 7 billion parameters from other companies on various benchmarks, including biology, physics, chemistry-related questions, code generation, math problems, and reasoning tests.

Meta llama 3 Pre-trained model performance

The Llama 3 70B model is also very good at competing with some of the best AI models on the market. In many ways, it’s better than Google’s Gemini 1.5 Pro, and in some scores, it’s even better than some models in the Claude 3 series

Meta Llama 3 instruct model performance

Meta has also made its own tests, which cover everything from code to creative writing. Although Llama 3 70B excels in these tests, it’s important to approach these results cautiously, given that Meta is the one who created them.

Meta Llama 3 Instruct Human evaluation

On a more qualitative level, Meta highlights that the new Llama models offer better control, are less likely to refuse answering questions, and show improved accuracy in trivia, history, STEM subjects, and coding recommendations. This marks a considerable step forward in making AI tools more reliable and versatile for users.

Meta’s Llama 3 training data

Meta has been tight-lipped about the origins of its training data for the new Llama 3 AI models, sharing only that it used “publicly available sources.” The company improved its dataset significantly by including four times more coding data than the previous Llama 2 dataset and incorporating 5% non-English content from around 30 different languages. This is aimed at boosting the model’s ability to handle multiple languages effectively.

Meta has also used “synthetic data” which is data made by Llama 2 AI systems, to create longer text documents that will be used to teach Llama 3 models. This approach is somewhat controversial because it may result in performance issues for the AI. Meta says that the large amount of different data has helped the models get better at understanding minor differences in language and doing well in many different tasks, as explained in a blog post. Like many AI companies, Meta views its training data as a vital asset and keeps details under wraps partly to avoid potential legal issues related to intellectual property.

New features for the Llama 3 series on the way

The company is currently working on bigger models that can handle more than 400 billion factors. These models are made to improve how users interact with them by including features like the ability to process different types of media and multilingual support for global communication. Meta plans to release these advanced models in the coming months. They will have a lot more features and a bigger context window for operations.

Meta Lama 4 40B+

Even though these new models are still in the training phase, Meta has shared some early results that show how useful they could be. Based on early checkpoints, these insights point to promising advances in AI capabilities, even though they are not yet incorporated into the models that are available. Meta will publish a detailed research paper that gives a full overview of Llama 3’s features and performance improvements once the training is over. 

Where to get your hands on Meta Llama 3

Meta has now incorporated its latest Llama 3 technology into Meta AI, positioning it as a leading AI assistant globally. You can now use this cutting-edge technology on Facebook, Instagram, WhatsApp, Messenger, and the web, among other places. Meta AI can assist people get things done, learn new things, create creative content, and connect with important aspects of their lives. Visit Meta AI’s official page to learn more about the experience in more depth.

If you want to use these models, the Llama 3 website has downloads and a full Getting Started Guide that lists all the sites where Llama 3 can be used. It will also soon be possible for users of Ray-Ban Meta smart glasses to experience multimodal Meta AI features. 

In short 

Meta has released the latest iterations of its Llama AI series, including the Llama 3 8B and 70B models which boast cutting-edge capabilities. These open-source models are designed to provide developers around the world with more sophisticated tools for AI application development. With up to 70 billion parameters, these models should provide better performance, making tasks like text processing, reasoning, and coding better. As part of its dedication to innovation, Meta wants to make these models even better by adding support for more languages and modalities. This will make them much more useful in a lot more fields. 

Related terms

Related articles

About XPS's Editorial Process

XPS's editorial policy focuses on providing content that is meticulously researched, precise, and impartial. We adhere to rigorous sourcing guidelines, and every page is subject to an exhaustive review by our team of leading technology specialists and experienced editors. This method guarantees the integrity, pertinence, and utility of our content for our audience.

Maryan Duritan
Maryan Duritan
Maryan Duritan, a seasoned U.S.-based copywriter and SEO specialist, excels in making complex ideas accessible. She crafts compelling website content, blogs, articles, ebooks, press releases, and newsletters, tailoring tone and voice to match client goals and audience needs. Her creative precision transforms ideas into impactful content.

Why Trust Us

Our editorial policy emphasizes accuracy, relevance, and impartiality, with content crafted by experts and rigorously reviewed by seasoned editors for top-notch reporting and publishing standards.

Disclosure
Purchases via our affiliate links may earn us a commission at no extra cost to you, and by using this site, you agree to our terms and privacy policy.

Latest articles

Most popular

Latest articles

Popular categories

Artificial intelligence

Artificial Intelligence (AI) is a branch of computer science focused on creating systems that emulate human intelligence.

Cryptocurrency

Cryptocurrency is a digital currency secured by cryptography and operates without a central authority.

Tech

Latest developments in technology, including new gadgets, software updates, industry trends, and breakthroughs in science and innovation.

Gaming

Covers updates and developments in the video game industry, including new game releases, updates, reviews, and events.

Cybersecurity

Cybersecurity is protecting computer systems and data from digital attacks and unauthorized access.

Investing

Provides updates on financial markets, stock performances, economic trends, and investment strategies.

VPN

Updates on VPN technology, security features, service providers, privacy issues, and changes in regulations affecting VPN usage.

Networking

Networking connects computers and devices to share resources and information using hardware and software.