Meta introduces a new, more efficient Llama model.

Meta has introduced the latest addition to its family of generative artificial intelligence models, the Llama 3.3 70B. In an announcement on X, Ahmad Al-Dahle, Vice President of Generative AI at Meta, highlighted that this text-only model matches the performance of Meta’s largest model, the Llama 3.1 405B, but at a lower cost. Al-Dahle noted that recent advancements in post-training techniques have significantly improved the overall performance of the model.

A graph was shared showing the Llama 3.3 70B outperforming competitor models such as Google’s Gemini 1.5 Pro, OpenAI’s GPT-4o, and Amazon’s Nova Pro across various industry benchmarks, including MMLU, which measures a model's ability to understand language. A Meta spokesperson communicated that this model should provide enhancements in areas such as mathematics, general knowledge, instruction-following, and application usage.

The Llama 3.3 70B is available for download on the AI development platform Hugging Face and other sites, including the official Llama website. This is part of Meta’s strategy to consolidate its position in the field of artificial intelligence by offering "open" models that can be used and commercialized across various applications. However, the usage terms set by Meta restrict how some developers can utilize the Llama models; platforms with more than 700 million monthly users must apply for a special license. Nevertheless, this restriction has not prevented the Llama models from accumulating over 650 million downloads.

Meta has also been using Llama internally; its AI assistant, Meta AI, is fully powered by Llama models and boasts nearly 600 million monthly active users, according to Meta CEO Mark Zuckerberg. The latter has claimed that Meta AI is on track to become the most widely used AI assistant globally.

For Meta, the open nature of Llama has been both an advantage and a disadvantage. Recently, it was alleged that Chinese military researchers used a Llama model to develop a defensive chatbot, prompting Meta to make its models available to defense contractors in the U.S. Additionally, the company has expressed concerns about its ability to comply with the EU AI Act, which establishes a regulatory framework for AI, describing the implementation of the law as "even unpredictable" for its open launch strategy.

A related issue involves the provisions of the GDPR, the EU privacy law, concerning AI training. Meta trains its AI models using public data from Instagram and Facebook users who have not opted out, data that in Europe is subject to GDPR safeguards. Earlier this year, EU regulators asked Meta to suspend training with data from European users while assessing the company's compliance with GDPR. Meta complied with the request while backing an open letter calling for "a modern interpretation" of GDPR that does not "stifle progress."

Moreover, Meta faces technical challenges similar to those encountered by other AI labs and is expanding its computing infrastructure to train and serve future generations of Llama. Recently, the company announced it would build a $10 billion AI data center in Louisiana, the largest Meta has constructed to date. Zuckerberg mentioned in the August fourth-quarter earnings call that, to train the next major set of Llama models, Llama 4, the company will need ten times more computing capacity than was used to train Llama 3. Meta has secured a cluster of over 100,000 Nvidia GPUs for model development, competing with resources from rivals like xAI. Training generative AI models is a costly operation, and Meta's capital expenditures rose nearly 33%, reaching $8.5 billion in the second quarter of 2024, compared to $6.4 billion the previous year, driven by investments in servers, data centers, and network infrastructure.