Meta Launches Llama 4: Powerful Multimodal AI with 10M Token Context Window
On April 5, 2025, Meta unveiled Llama 4, its latest open-source AI series, introducing two standout models: Llama 4 Scout and Llama 4 Maverick.
Scout, with 17 billion parameters and a massive 10 million token context window, outperforms models like Gemma 3 and Mistral 3.1.
Maverick is built for multimodal tasks and surpasses GPT-4o and Gemini 2.0 Flash, scoring 1417 ELO on the LMArena benchmark.
Both models leverage a mixture-of-experts architecture and iRoPE technology to boost performance and efficiency.
Meta is also training a larger model, Llama 4 Behemoth, which already outperforms GPT-4.5 on STEM benchmarks—even before completion.
The models are available on Hugging Face, though their usage is currently restricted in the EU due to strict privacy regulations.