Meta presented Llama 4, its most recent series of artificial intelligence models. The models are now the basis of the assistant Meta Ai, available on the web and integrated in various messaging platforms such as WhatsApp, Messenger and Instagram. Llama 4 Scout is a compact model, able to operate on a single NVIDIA H100 GPU. Llama 4 Maverick, on the other hand, is similar to GPT-4o and Gemini 2.0 Flash. Meta is still conducting the training of Llama 4 Behemoth, which CEO Mark Zuckerberg defines the basic model with the highest performance at a global level.
Lama 4 Scout boasts a context window of 10 million token, comparable to the working memory of a complete model of artificial intelligence. In the main benchmarks it exceeds the Gemma 3 and Gemini 2.0 flash-lite models of Google, as well as the Open-Source Mistral 3.1 model. The ability of Llama 4 Scouts to operate on a single NVIDIA H100 GPU is highlighted. Meta formula similar statements regarding the performance of its largest model, Maverick, compared to GPT-4o by Openi and Gemini 2.0 Flash of Google. The company claims that the results obtained are comparable to Deepseek-V3 in coding and reasoning tasks, using a significantly lower number of active parameters.
Llama 4 Behemoth presents an imposing architecture with 288 billion active parameters and a total of 2 trillions of parameters. Although it has not yet been officially released, Meta claims that Behemoth is able to overcome its competitors, such as GPT-4.5 and Claude Sonnet 3.7, in various benchmark STEMs. In contrast, Llama 4 uses “Mixth of Experts” type architecture (MOE), an approach that optimizes the use of resources by using only the necessary components of the model for a specific task. Meta plans to reveal its future plans for the models and artificial intelligence products during its next conference Llamacon, scheduled for April 29. Similarly to previous models, Meta classifies the Llama 4 series as “open-source”. However, Llama received criticism for its restrictive terms. For example, the Llama 4 license imposes on commercial entities with over 700 million active users monthly to obtain the authorization from a destination before being able to use its models. This license requirement, as highlighted by the Open Source Initiative in 2023, led to its exclusion from the “open source” category in the strict sense.