Meta has unveiled Llama 3.3 70B, the latest addition to its Llama series of generative AI models. Ahmad Al-Dahle, Meta's VP of generative AI, announced on X that this new text-only model matches the performance of the larger Llama 3.1 405B model but at a reduced cost. According to Al-Dahle, advancements in post-training techniques allow Llama 3.3 70B to deliver superior performance while being more efficient.
Meta shared benchmark results showing Llama 3.3 70B surpassing Google’s Gemini 1.5 Pro, OpenAI’s GPT-4o, and Amazon’s Nova Pro in areas like language understanding, as measured by tests such as MMLU. The model offers enhancements in math, general knowledge, instruction following, and app usage. It is available for download on platforms like Hugging Face and the official Llama website.
While Meta promotes Llama as “open,” its terms restrict certain use cases, such as requiring special licenses for platforms with over 700 million monthly users. Despite these constraints, Llama models have been downloaded more than 650 million times, and Meta AI—powered entirely by Llama—boasts nearly 600 million active monthly users. Meta CEO Mark Zuckerberg predicts Meta AI could become the world’s most widely used AI assistant.
However, Llama’s open nature has sparked challenges. A report revealed Chinese military researchers used Llama to create a defense chatbot, prompting Meta to make its models available to U.S. defense contractors. Additionally, Meta has faced scrutiny under the EU’s AI Act and GDPR regulations. In response to privacy concerns, EU regulators requested Meta halt AI training on European user data, leading the company to advocate for a more flexible interpretation of GDPR.
Meta is also scaling its infrastructure to prepare for future Llama generations, including Llama 4, which will require 10 times more computational power than Llama 3. To meet this demand, Meta is investing $10 billion in a new AI data center in Louisiana, its largest to date, and has acquired over 100,000 Nvidia GPUs for model training.
These investments come at a steep cost. Meta’s capital expenditures rose 33% in Q2 2024, reaching $8.5 billion, driven by the need for enhanced servers, data centers, and network capabilities.
Post a Comment