Llama 4: Meta launched the new generation multimodal AI models with three versions

Target (5) The target announced on Saturday, Lama 4, the new generation of the large -scale language model (LLM), which comes with three versions still being developed, is still being developed It has been started more powerfully by the company so far. Each focuses on different tasks.

Lama 4 Scout, one of the premiere, uses a combination of 17 billion active parameters (MOE) and was trained with 16 experts. According to Big Teh, it offers a context window of 10 million jets, can be used with the NVIDIA H100 GPU and exceeds models like Gemma 3 and Gemini 2.0 flash Lite and Mistral 3.1 from Google.

Lama 4 Maverick brings the same 17 billion active parameters, but there were 128 training experts. The owner of Facebook has half of the active parameters, with the comparable results with Deepek V3 in coding and reasoning, indicating that the GPT-4O and Gemini 2.0 flash in a wide range of accepted criteria.

Lama 4 Behemoth. Still in education, he got it GPT-4.5, Claude Sonnet 3.7 and Gemini 2.0 PRO higher results focused on STEMAs the technology giant explains.

What is the Multimodal AI model?

Multimodal AI Systems Working ability with different data typesCombining and analyzing various forms of input with a wider understanding produces solid results. In contrast, traditional models are only interested in one type.

For example, when sharing a photo with a multimodal version, it can create a text summary that explains what you see in the image and can do the opposite, analyze a text request by creating a desire to create an image. Video and sound are other compatible formats depending on the version.

In the new generation target models, Lama 4 Scout is an option dedicated to users who want to work with texts and images as described by Mark Zuckerberg. This is considered a more fundamental version, this serves professionals and companies looking for smart tools for less complex tasks.

The more advanced Lama 4 Maverick has the capacity to operate with higher performance activities such as large amounts of information processing for summaries of various documents suitable for medium and large corporate users. Lama 4 Behemoth will be designed for companies that want to develop their own models.

The target there with Lama 4

According to the technology giant, the capabilities of the new generation Llama 4 models can be tested on the AI ​​target in the web version of WhatsApp, Messenger, Instagram and platform. The update will be launched in more than 40 countries and in 13 languages.

However, new multimodal features Currently limited to the United States and English language. The target said that without confirming a more accurate deadline date, Lama 4 will do AI for more people around the world later this year.

Open Source, Lama 4 Scout and Lama 4 Maverick can also be downloaded from the model website and hugging facial platform and allow developers to add new tools to their products.

Do you like the content? Don’t forget to share this and other news from Tecmundo on social networks.