Meta推出Llama 4开源模型，在代码生成和数学推理上超越GPT-4。
Meta released Llama 4 today, the latest version of its open-source large language model. With 70 billion parameters, Llama 4 achieves 92% accuracy on the HumanEval code generation benchmark and 95% on GSM8K math reasoning, both surpassing GPT-4. Meta highlighted that Llama 4 uses a novel sparse attention mechanism, improving inference efficiency by 40%. The model is released open-source with commercial use allowed, expected to drive innovation in AI communities and enterprise applications. Llama 4's release also intensifies competition between open-source and closed-source models.