Meta: Llama Guard 4 12B

Meta's Llama Guard 4 12B, a multimodal model derived from Llama 4 Scout and fine-tuned for content safety classification, demonstrates exceptional performance in operational efficiency. It consistently ranks among the fastest models, achieving an Infinityth percentile across five benchmarks, indicating unparalleled speed. Similarly, its pricing is highly competitive, also securing an Infinityth percentile across five benchmarks. However, the model's performance on the provided benchmark results for General Knowledge, Ethics, Email Classification, Instruction Following, and Coding is uniformly 0.0% accuracy. This suggests that while Llama Guard 4 excels in speed and cost-effectiveness, its current iteration, as evaluated by these specific benchmarks, does not demonstrate proficiency in general cognitive tasks, ethical reasoning, classification, instruction adherence, or coding. Its core strength lies in its intended purpose as a content safety classifier, leveraging multimodal capabilities for English and multiple languages, including handling mixed text-and-image prompts. The benchmark results likely reflect its specialized design for safety classification rather than general-purpose AI tasks. Its integration into the Llama Moderations API further underscores its role as a dedicated safety tool.