Meta dropped Llama 4: What do you know about the new AI’s models

Mita revealed its latest members from Obstetric artificial intelligence model on saturday. You can give Lama 4 Test drivers now through Meta AI, Llama 4 will soon run a lot Amnesty International dead Features on the services of the company, WhatsApp and Messenger.

The competition between Meta and other artificial intelligence companies has become increasingly intense. Companies are working to build and issue Amnesty International models capable of more complicated tasks and advanced thinking without the need for huge amounts of computing power and operation. It is a difficult sweet spot for beating, and Meta hopes to put it in its latest models at the top of the competitors like Chatgpt and twin.

There are models in the Llama 4 family now available: Scout and MAVERICK. They are open weight models, which means that they can create text, photos and symbols. Open models like Meta means that developers can get an insight into how to build models. Llama 4 models are open weight models, which means that you can learn how the model makes connections and how certain properties are given larger than weight as it learns. Openai announced earlier this month that it is developing a file Open weights model For the first time.

Scout is the smallest family model, designed to run on one NVIDIA H100 graphics unit. Scout has a symbolic context window of 10 million, a parameter model 17 billion that includes 16 experts (sub -networks within the model, allowing them to operate tasks more efficiently). Scouts have more than twice the power of fire Lama 3Which has 8 billion teachers. In general, the greater the number of parameters owned by the model, the greater its ability to provide better results faster.

MAVERICK is a medium -sized model, The Big Brother to Scout, and includes 17 billion teachers with 128 experts. Meta said that the standard tests showed the MAVERICK Beat Chatgpt-4O and the generation of texts. Dibsic v3. Dibsic still has the edge when it comes to logic and coding. CNET has not independently realized of Meta measurement tests.

More information is expected to come about the rest of the Llama 4 family, including a basic model called Behymoth and Llama 4 logical model, later this month, According to a video published by CEO Mark Zuckerberg. We will likely get to know more about these models in Llamacon, the first annual AI’s annual conference in the company starting on April 29.

To find out more, check what we know about a A possible independent application for Meta AI And our review of Best Chatbots Amnesty International.