Meta has officially confirmed it is in the process of training its next-generation large language model, Llama 3.1, a significantly more powerful version designed to compete directly with the industry’s most advanced models, including OpenAI’s GPT-4o and Google’s Gemini 1.5 Pro. The announcement signals Meta’s ambition to not only lead in the open-source community but to achieve state-of-the-art performance on a global scale.
According to a blog post from the AI@Meta team, Llama 3.1 is a 405B parameter model, a massive leap from the 70B parameter version of Llama 3 released earlier this year. This new model is being trained on Meta’s custom-built AI infrastructure, which consists of two 24,000-GPU clusters. The company stated that Llama 3.1 is demonstrating significant improvements in key areas such as complex reasoning, nuanced code generation, and advanced instruction following.
While Llama 3 was released as an open-source model, Meta has not yet confirmed the release strategy for the 405B version of Llama 3.1, though a series of smaller, more accessible versions are expected to be made available soon. The primary goal of the flagship model is to power a new wave of AI features across Meta’s product ecosystem, including more sophisticated AI assistants for Instagram, WhatsApp, and Messenger, as well as enhancing the capabilities of its Ray-Ban smart glasses.
The development of Llama 3.1 underscores the relentless pace of the AI arms race. By building a model that aims to surpass the performance of its rivals, Meta is reinforcing its position as a key player in the foundational model space. The success and potential release of such a powerful model could provide developers and businesses with an open-source alternative that rivals the best proprietary systems, further accelerating innovation across the industry.
English 
Deutsch