Meta, the company behind Facebook and Instagram, has just introduced a new version of its artificial intelligence (AI) model called Llama 3.3 70B. This model is part of the Llama family, which is known for generating text. The key feature of Llama 3.3 70B is that it performs almost as well as Meta’s previous largest model, Llama 3.1 405B, but at a much lower cost.
Ahmad Al-Dahle, who is the Vice President of generative AI at Meta, shared this news on a social media platform called X. He explained that the new model uses the latest techniques to improve its performance while keeping expenses down. In simple terms, they’ve found a way to make the AI smarter without spending as much money.
Al-Dahle also provided a comparison chart showing that Llama 3.3 70B outperforms several other well-known AI models, including Google’s Gemini 1.5 Pro, OpenAI’s GPT-4o, and Amazon’s Nova Pro. These comparisons were based on various tests that measure how well the AI understands language and performs tasks. According to a representative from Meta, this new model is expected to be better at tasks like math, general knowledge, following instructions, and using apps.
The Llama 3.3 70B model is available for download from several platforms, including Hugging Face and the official Llama website. Meta aims to lead the AI industry by providing these “open” models, which means they can be used by developers for various applications, including commercial ones. However, there are some restrictions on how certain developers can use these models. For example, platforms with over 700 million monthly users need to get special permission to use Llama models. Despite these limitations, Llama models have been downloaded over 650 million times, showing their popularity.
Read Also: Grok Is Now Free For All X Users
Internally, Meta has been using Llama models as well. Their AI assistant, Meta AI, which relies entirely on Llama technology, has around 600 million active users each month. Meta CEO Mark Zuckerberg has claimed that Meta AI is on track to become the most widely used AI assistant globally.
While the open nature of Llama has its advantages, it has also led to some challenges for Meta. Recently, there were reports that researchers linked to the Chinese military had used a Llama model to create a defense chatbot. In response, Meta decided to make its Llama models available to U.S. defense contractors.
Meta has also expressed concerns about complying with the AI Act, a new law in the European Union (EU) that regulates AI technologies. They find the law’s implementation to be unpredictable, which complicates their strategy of releasing their models openly. Additionally, there are issues related to the General Data Protection Regulation (GDPR), which governs how personal data is used in the EU. Meta trains its AI models using public data from Instagram and Facebook users who haven’t opted out, but this data is subject to GDPR rules. Earlier this year, EU regulators asked Meta to stop using European user data for training while they checked if the company was following GDPR guidelines. Meta agreed to this request while also supporting a call for a modern interpretation of GDPR that would allow for more innovation.
Meta is also facing the same technical challenges as other AI companies and is working to improve its computing resources to train future Llama models. Recently, the company announced plans to build a massive AI data center in Louisiana, which will cost $10 billion and will be the largest of its kind that Meta has ever constructed. Zuckerberg mentioned in a recent earnings call that to develop the next version of Llama, called Llama 4, they will need ten times the computing power that was required for Llama 3. To achieve this, Meta has acquired a large number of advanced graphics processing units (GPUs) from Nvidia, which puts them on par with other major players in the AI field.
Training AI models is an expensive endeavor. Meta’s spending on capital projects has increased significantly, rising nearly 33% to $8.5 billion in the second quarter of 2024, compared to $6.4 billion a year earlier. This increase is primarily due to investments in servers, data centers, and network infrastructure needed for their AI initiatives.
