Meta has released what it’s calling a new “herd” of Llama 4 models. There are three flavors of the new Llama 4, called Scout, Maverick and Behemoth, and two are available right now for you to try in your Meta apps like Messenger, Instagram and WhatsApp.
Llama 4 is the latest flagship version of Meta’s open source Llama AI, and the new release comes almost exactly a year after the release of Llama 3 in 2024.
Inspired by the training advancements made by DeepSeek, the new Llama 4 has been trained using the more efficient ‘mixture of experts’ methodology.
As the names suggest, Scout is the most lightweight model, with 109 billion parameters, while Maverick has 400 billion parameters. Both of these models are available right now for developers to download, and are also used in the popular Meta consumer apps.
Education-heavy
Llama 4 Behemoth is a teacher-focused model, which Meta claims out performs GPT-4.5, Claude, Sonnet 3.7 and Gemini 2.0 Pro on STEM-focused benchmarks such as MATH-500 and GPQA Diamond. Currently there is no access to Llama 4 Behemoth, as Meta says it is still “in training”.
Meta’s new models keep it at the forefront of competitive open source LLMs. While the benchmarks are impressive, the current consumer experience of using Meta AI still lags far behind using ChatGPT or Gemini.
For example, while both the two available Llama 4 AIs are multimodal, there is still no way to upload an image via meta.ai, or in one of the many Meta apps. You can ask Meta AI to look at the URL of an image and analyze what is sees, but direct upload isn’t supported.
Equally, Meta AI lacks other chatbot extras we’ve come to think of as standard these days, like AI search and deep reasoning, and its image generation capabilities lag behind the most recent ChatGPT update.
Copyright issues remain
The new Llama 4 models are accessible to developers, who can download the open source models to use at competitive token rates at llama.com and Hugging Face.
Alternatively the new Llama 4 LLMs are available right now to use at Meta.ai or in the Meta apps like Messenger, WhatsApp and Instagram Direct.
It’s worth noting that the new Llama 4 LLMs remain part of an ongoing copyright dispute between Meta and several famous authors after court documents alleged that Meta CEO Mark Zuckerberg had approved the use of the LibGen data set, amongst other shadow libraries, in training its Llama LLM.
The Atlantic recently published a searchable database of titles contained in LibGen, enabling many authors to see if Meta could have been training its AI on their work without permission.