Meta Introduces Llama 3.3 70B: A More Efficient AI Model

Meta has expanded its Llama family of generative AI models with the release of Llama 3.3 70B. This new text-only model offers the same level of performance as Meta’s earlier Llama 3.1 405B but at a significantly reduced cost.

Ahmad Al-Dahle, Meta's Vice President of Generative AI, shared in a post on X that advancements in post-training techniques have enabled Llama 3.3 70B to improve core performance while being more cost-effective. A chart provided by Al-Dahle showed the model outperforming competitors like Google’s Gemini 1.5 Pro, OpenAI’s GPT-4o, and Amazon’s Nova Pro on benchmarks such as MMLU, which evaluates language understanding. According to a Meta spokesperson, the new model excels in tasks such as mathematics, general knowledge, instruction following, and app integration.

Llama 3.3 70B is available for download on platforms like Hugging Face and Meta’s official Llama website. Although not entirely “open source,” the model is accessible for various commercial applications. However, developers operating platforms with over 700 million monthly users are required to obtain a special license. Despite this limitation, Llama models have seen over 650 million downloads.

Internal and External Uses of Llama Models

Meta uses Llama models extensively within its operations. For instance, Meta AI, the company’s assistant powered by Llama, serves nearly 600 million monthly active users. CEO Mark Zuckerberg has stated that Meta AI is on track to become the most widely used AI assistant globally.
However, the model’s open accessibility has posed challenges. In one case, Chinese military researchers reportedly used a Llama model to develop a defense chatbot. In response, Meta made its Llama models available to U.S. defense contractors.
The company has also faced regulatory concerns in Europe. The EU’s AI Act and GDPR provisions have prompted Meta to temporarily halt training on European user data to ensure compliance. Meta continues to advocate for updated regulations that balance innovation with privacy protections.

Investing in AI Infrastructure

To support the development of future models, Meta is ramping up its computing infrastructure. The company recently announced a $10 billion investment in a new AI data center in Louisiana, its largest to date. Training the next iteration of Llama models, such as Llama 4, will require ten times the computing power used for Llama 3, according to Zuckerberg. To meet this demand, Meta has secured over 100,000 Nvidia GPUs, placing its resources on par with major competitors.

Rising Costs in Generative AI Development

Building and training generative AI models is an expensive endeavor. Meta's capital expenditures increased by 33% in Q2 2024, reaching $8.5 billion, compared to $6.4 billion the previous year. These costs are driven by investments in servers, data centers, and network infrastructure necessary for advancing its AI capabilities.
Llama 3.3 70B represents Meta’s ongoing commitment to innovation in generative AI, balancing performance improvements with cost-efficiency while navigating regulatory and operational challenges.