Over the weekend, Meta introduced three large language models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth, which are collectively referred to as “the world's smartest large language models.”
In a blog post, Meta revealed that Llama 4 Scout is a general-purpose model with 17 billion active parameters and 109 billion total parameters. This compact model has 16 experts, a context window of 10 million tokens, and fits on a single NVIDIA H100 CPU. The large language model (LLM) is said to outperform Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 on a broad range of benchmarks.
With a total of 400 billion parameters and 17 billion active parameters per command, Meta calls Llama 4 Maverick “the best multimodal model in its class.” It has a context window of 1 million tokens and uses 128 routed experts.
Like Llama 4 Scout, this new language model uses a mixture of experts (MoE) architecture, meaning that a single token activates only a fraction of the total parameters. Therefore, according to Meta, LLMs with MoE architectures are more compute-efficient and deliver higher quality compared to a dense model.
The tech company claims that Llama 4 Maverick performs better than GPT-4o and Gemini 2.0 Flash in certain benchmarks and performs about as well as DeepSeek v3 in coding and reasoning while having about half as many active parameters.
Both Llama 4 Scout and Llama 4 Maverick are distillations from Llama 4 Behemoth. This multimodal model counts nearly 2 trillion total parameters, including 288 active billion parameters per command. In addition, Llama 4 Behemoth has 16 experts and is Meta’s most powerful language model ever.
Meta says Llama 4 Behemoth is its “most powerful [model] yet and among the world’s smartest LLMs,” outperforming ChatGPT 4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro in benchmarks such as LiveCodeBench, MATH-500, and GPQA Diamond.
Llama 4 Behemoth is still being trained, so Meta has decided not to make it available to the public yet. Llama 4 Scout and Llama 4 Maverick are available for download on llama.com and Hugging Face. Users can also try these models in WhatsApp, Messenger, Instagram Direct, and on the Meta.AI website.
Meta will provide more details on the Llama 4 family at LlamaCon, a developer conference that will take place on April 29th.
Your email address will not be published. Required fields are markedmarked