In order to compete with OpenAI‘s open-source GPT-4, which drives products like ChatGPT and Microsoft Bing, Meta announced that it is opening up the source code for their large language model, LLaMA 2. This makes it free for commercial and research use.
As part of Microsoft’s Inspire event, Meta made the announcement, highlighting its support for Windows and Azure as well as the two companies “growing” cooperation. At the same time, Microsoft provided additional information regarding the cost and features of the AI capabilities included in its 360 platform. Additionally, Qualcomm disclosed that it is collaborating with Meta to introduce LLaMa to laptops, phones, and headsets starting in 2024 for AI-powered apps that don’t require the use of cloud services.
According to Meta’s press release, the decision to open up LLaMA was made in order to provide companies, startups, and researchers with wider access to AI technologies and enable community experimentation. LLaMa 2 reportedly received training on 40% more data than LLaMa 1, which includes data from “publicly available online data sources.” Additionally, it asserts that it “outperforms” other LLMs like Falcon and MPT on assessments of thinking, coding, competency, and understanding.
Meta stated that by opening up LLaMA, it hopes to increase transparency and safety. According to the company, the LLaMA 2 model has been “red-teamed,” or evaluated for safety by “generating adversarial prompts to facilitate model fine-tuning,” both internally and externally. Meta also reveals the evaluation and modification processes for the models. The Microsoft Azure platform will offer the open-sourced LLaMa 2. According to Meta, LLaMA will also be offered via Hugging Face, AWS, and other service providers.
Meta revealed in a statement that “We believe an open approach is the right one for the development of today’s AI models, especially those in the generative space where the technology is rapidly advancing. Opening access to today’s AI models means a generation of developers and researchers can stress test them, identifying and solving problems fast, as a community.”
Meta’s LLaMA model first appeared in February, but a few days later, it was leaked on 4chan. With more individuals having access to LLaMA 2, more AI-powered solutions will undoubtedly be developed based on the model. Although Meta claims to have had more than 100,000 requests from researchers to utilize its initial model, it is expected that the open-source LLaMA 2 will be extensively used.