Meta unveils Llama 4 – new contender in the AI race

MENLO PARK, CALIFORNIA: Meta Platforms Inc. has unveiled its latest Llama 4 artificial intelligence models, which it claims are among the most advanced large language models globally. At the heart of the Llama 4 collection is Llama 4 Behemoth, a two-trillion parameter language model that is still under development.

However, two derived models, named Maverick and Scout, are immediately available for developers and users to explore through Meta’s applications and the Meta.ai website. These models are touted as natively multimodal, meaning they can handle various types of data beyond just text, and Meta asserts that their benchmark scores surpass those of well-known competitors.

Meta’s announcement comes amid a competitive push from U.S. companies to respond to a wave of new AI models and services introduced by Chinese tech firms in recent weeks. Following the launch of DeepSeek in January, companies like Alibaba Group Holding Limited and Tencent Holding Limited have accelerated their AI initiatives, enhancing investments to secure a strong market position. For instance, Baidu Inc. has made its Ernie Bot freely available, Tencent has integrated DeepSeek into its widely-used WeChat app, and Alibaba has introduced several models claiming to outperform DeepSeek.

Italy’s PM Meloni meets Meta exec to discuss investments

Based in Menlo Park, California, Meta is positioning itself in competition with DeepSeek, as well as OpenAI and Alphabet Inc.’s Google, while also addressing the cost implications for users of its new Llama tools. For the first time, the company is implementing a Mixture of Experts (MoE) architecture, which DeepSeek successfully utilized to optimize model costs. This approach aims to improve efficiency by breaking down the workload into smaller, more manageable components.