Facebook's parent company has just announced two new models in the Llama 4 collection. This is the same model that is being used for the Meta AI assistant on the web as well as in WhatsApp, Messenger and Instagram, helping people create more personalized multimedia experiences.
The two new models, available for download now from Meta or Hugging Face, are the Llama 4 Scout, and the Llama 4 Maverick. Meta says it's still in the process of training the Llama 4 Behemoth, a massive model that CEO Mark Zuckerberg says is “the world's highest-performing platform model.”
According to Meta, Llama 4 Scout has a context window of up to 10 million tokens, many times more than the current model with the highest number of tokens, Gemini. This number represents the working memory of an AI model.
Meta describes the model as outperforming Google's Gemma 3 and Gemini 2.0 Flash-Lite, as well as Mistral 3.1 “on many popular publicly available benchmarks,” while still being able to “run on a single Nvidia H100 GPU.”
For Llama 4 Maverick, the company makes similar performance claims when compared to OpenAI's GPT-4o and Google's Gemini 2.0 Flash. Maverick's programming and inference tasks are comparable to DeepSeek-V3, even using less than half the parameters.
 |
Figures of the new models in the Llama 4 collection. Photo: Meta. |
Behemoth, meanwhile, has 288 billion active parameters out of a total of 2,000 billion. Although it has not yet been released, Meta says Behemoth is capable of outperforming its competitors (specifically GPT-4.5 and Claude Sonnet 3.7) “on a number of STEM metrics.”
Meta said it has switched to a “mixture of experts” (MoE) architecture, which has been a success for DeepSeek. The approach saves resources by only activating the parts of the model that are needed for a particular task. More details on the product plans will be announced at the LlamaCon conference, which takes place on April 29.
Despite criticism over its licensing restrictions, Meta still calls Llama 4 “open source,” as it has done with previous models. For example, Llama 4’s license requires commercial organizations with more than 700 million monthly active users to seek Meta’s permission before using the model.
According to the Open Source Initiative in 2023, such a regulation “renders the model no longer open source.”
Source: https://znews.vn/cha-de-facebook-ra-mo-hinh-ai-moi-post1543744.html
![[Photo] Visiting Cu Chi Tunnels - a heroic underground feat](https://vstatic.vietnam.vn/vietnam/resource/IMAGE/2025/4/8/06cb489403514b878768dd7262daba0b)
Comment (0)