Meta has launched the latest addition to its Llama family of generative AI models, the Llama 3.3 70B, promising to deliver top-tier performance at a reduced cost. The model, which is text-only, offers similar capabilities to Meta’s largest Llama model, Llama 3.1 405B, but with a more efficient use of resources.
In a post on X, Ahmad Al-Dahle, Meta’s Vice President of Generative AI, explained that Llama 3.3 70B incorporates cutting-edge post-training techniques to enhance its core performance while significantly lowering costs. A chart shared by Al-Dahle showed Llama 3.3 outperforming competitors, including Google’s Gemini 1.5 Pro, OpenAI’s GPT-4, and Amazon’s Nova Pro, on several industry benchmarks, including MMLU, which tests language understanding.
Meta claims the new model improves performance in key areas such as mathematics, general knowledge, instruction-following, and app interaction and available for download through platforms like Hugging Face and Meta’s official Llama website.
Llama 3.3 70B is part of Meta’s strategy to dominate the AI space with “open” models that can be leveraged for a wide range of commercial applications. While the models are accessible to many developers, certain restrictions apply. Platforms with more than 700 million monthly users must obtain special licenses to use Llama models, which presetly has more than 650 million downloads, according to Meta.
Meta’s internal use of Llama has been significant as well. The company’s AI assistant, powered entirely by Llama models, now boasts nearly 600 million monthly active users. Meta CEO Mark Zuckerberg has claimed that Meta AI is on track to become the world’s most-used AI assistant.
In November, reports surfaced that Chinese military researchers had used a Llama model to develop a defense chatbot. In response, Meta made its models available to U.S. defense contractors. Meta has also expressed concerns about complying with the European Union’s AI Act and the General Data Protection Regulation (GDPR). EU regulators earlier this year requested that Meta halt training its AI on data from European users until they could assess compliance with GDPR rules.
Meta is also investing heavily in the infrastructure needed to support future generations of Llama models with the company recently announcing a $10 billion AI data center in Louisiana. Zuckerberg revealed that to train the next iteration, Llama 4, Meta will need ten times more computing power than required for Llama 3. The company has already secured a massive cluster of over 100,000 Nvidia GPUs to support this effort.
Training generative AI models is an expensive undertaking, and Meta’s capital expenditures have risen sharply as a result. In the second quarter of 2024, the company’s investments in servers, data centers, and network infrastructure grew by 33%, reaching $8.5 billion, up from $6.4 billion a year earlier.
As Meta pushes ahead with its ambitious AI initiatives, the company continues to navigate a complex landscape of regulatory, technical, and competitive challenges.
Philips Babatunde