Target has presented flame 4its new generation of artificial intelligence models, which comes to double the power and call processing capacity 3, the previous models.
The new family is composed of the new models called 4 Scout, Maverick and Behemoth and allows complex problems and communicating naturally as never before had the artificial intelligence of Meta. They have up to 288,000 million parameters and, according to the company, they are able to overcome the “smarter” models, such as Openai GPT-4.5.
The new generation of AI models 4, open sourceallows anyone to develop their experiences around them. According to Meta, call 4 scout and call 4 maverick are the first native multimodal models of open weight with unprecedented compatibility with the length of the context, “while” calls 4 Behemoth is one of the most intelligent models in the world and the most powerful to date, which will serve as a guide for the new models. “
Call 4 scout and call 4 maverick can already be downloaded from the website Llama.com and Hugging Face. They are also available for all users through WhatsApp, Messenger, Instant Messenger of Instagram Direct and the Meta AI website.
For its part, call 4 Behemoth is still in the development phase, so it is not available. Know now the characteristics of each of them:
Characteristics of the Meta Models 4
These are the characteristics of the different models of AI that are part of the generation calls 4 of goal and what each of them offers:
Call 4 scout
This is an AI model that is able to analyze the user’s activity to carry out personalized tasks, can reason about large code bases and is able to summarize a large number of documents, among other tasks. He is able to optimize his answers based on the nature of each question.
It has 17,000 million active parameters with 16 experts. The use of “experts” means that it only uses a group of total parameters for each task, which improves efficiency by reducing server and latency costs.
This model is capable of being installed in a single NVIDIA H100 GPU. Call 4 scout has a context window of ten million tokens, and according to Meta, it is exceeding models such as Gemma 3 and Gemini 2.0 Flash-Lite from Google and Mistral 3.1.
Call 4 maverick
It also has 17,000 million active parameters, but is based on 128 experts. As the previous one, it is also able to integrate into a single NVIDIA H100 DGX host, it offers results comparable with Depseek-V3 in coding and reasoning using less active parameters.
Its functions are the same as those of flame 4 scout, but in addition, it is able to understand and process both images and text. In addition, he would be able to create artificial intelligence applications.
With regard to its abilities, this model is also able to summarize, analyze and reason on code bases, but, unlike Scout, Maverick can understand both images and text and create sophisticated artificial intelligence applications.
Call 4 Behemoth
It is the most advanced of the models presented, although it is still in the training phase, in development, so it will not be available. Call 4 Behemoth has 288,000 million active parameters with 16 experts; And, according to Meta, it exceeds GPT-4.5, Claude Sonnet 3.7 and Gemini 2.0 Pro.
It offers avant -garde performance for models without reasoning in mathematics, multilingualism and images. When available, it will be used as a guide for new finishing models.
There is also a final model, he calls 4 reallocate that target has not yet provided much information.
