PANews reported on April 7 that Meta released its latest multimodal large language model Llama 4, including two versions: Scout and Maverick. Scout has 17B active parameters, 16 expert systems, and supports 10 million token context windows; Maverick has 128 expert systems and outperforms GPT-4o and Gemini 2.0 Flash in image understanding and cost-effectiveness. Meta also revealed that its training model Llama 4 Behemoth has surpassed GPT-4.5 and Claude 3.7 in the STEM benchmark.