Meta Partners With Microsoft to Commercialize the Next-Gen LLaMa 2 AI Model

Highlights:

• The LLaMA 2 family of large language models represent Meta’s next generative AI research iteration.
• As a preferred partner for the distribution of LLaMA 2, Microsoft Azure developers will have access to the AI models and cloud-native tools for training, fine-tuning, and deployment.

Meta Platforms Inc. recently announced that it has partnered with Microsoft Corp. to distribute its most recent artificial intelligence models for commercial use, enabling businesses to develop their AI-enabled apps and tools.

The LLaMA 2 large language model family represents Meta’s next generative AI research iteration. These AI models can also produce research, answer queries, compose poetry, generate computer code, comprehend human speech, and converse with humans.

Meta has a lengthy history of releasing its AI models as open-source software, and the entire LLaMA 2 family is currently available for free, including numerous pre-trained and fine-tuned models with 7 billion, 10 billion, and 70 billion parameters. According to Meta, LLaMA 2 was trained with over 40% more data than its predecessor and had an improved architecture; it was also fine-tuned with over 1 million human annotations to produce greater safety and quality.

Meta’s Chief Executive Mark Zuckerberg said, “Open source drives innovation because it enables many more developers to build with new technology. It also improves safety and security because when software is open, more people can scrutinize it to identify and fix potential issues.”

As a preferred partner for the distribution of LLaMA 2, Microsoft Azure developers will have access to the AI models and cloud-native tools for training, fine-tuning, and deployment. It is also possible to use its internal security features on Azure, such as Azure AI Content Safety, which can be combined with Meta’s filters to monitor content and provide a safer user experience.

Microsoft also stated that the AI models’ incorporation on Windows indicates that they are optimized for operation on that operating system. Developers will be able to operate it locally to create applications propelled by generative AI.

As the previous version of LLaMA was licensed only for research purposes, this news follows immediately after a report that Meta planned to release an AI model for commercial use. By entering the commercial market, Meta aims to alter the power dynamic between major participants, such as Microsoft-backed OpenAI LP, responsible for the popular ChatGPT and GPT-4, and Google LLC, which develops Google Bard.

After its parameters and weights were released to the public in March of this year, Meta’s original LLaMA model has become the foundation for numerous other open-source models. Its presence in the industry as open source also allows it to serve as the foundation for various projects, making it simpler to implement, and its commercial viability provides developers with easier access. “I believe it would unlock more progress if the ecosystem were more open, which is why we’re open-sourcing LLaMA 2,” added Zuckerberg.

Because OpenAI and Google’s AI models are so-called black boxes, or “closed source,” a Google engineer stated in May that the two companies have “no moat” in their respective industries. According to a previous report, Meta intends to keep the models free but may release a paid service for training and fine-tuning for enterprise customers in the future.

Meta is not just collaborating with Microsoft on LLaMA 2 support. Qualcomm Inc. announced recently that beginning in 2024, its smartphone chips and personal computers will support the new model. Qualcomm chips include a “tensor processor unit,” or TPU, which provides the types of computing capacity required for AI processing. Although LLMs typically require massive server farms to operate, these chips include a “tensor processor unit,” or TPU.

This will enable the smaller LLaMA 2 models to operate efficiently on smartphones and computers with these processors. If the AI operated on a phone or computer instead of transferring its computing power to a cloud compute farm, it would reduce the user’s privacy concerns and the cost of developing AI-enabled applications.