最新报道:Meta's Llama 4, released in April 2025, features three open-weight multimodal models: Scout (17B active parameters, 10M token context), Maverick (17B active parameters, 1M token context), and the upcoming Behemoth (288B active parameters). Available through cloud partners like AWS and Hugging Face, these models support text, image, and video tasks across 200 languages while incorporating safety tools like Llama Guard and Code Shield. However, limitations include potential code vulnerabilities (Maverick scored 40% on LiveCodeBench vs GPT-5's 85%) and training on controversial datasets.