Nous: Hermes 4 70B

Nous: Hermes 4 70B, a hybrid reasoning model built on Meta-Llama-3.1-70B, demonstrates a balanced performance profile with notable strengths in reliability and cost-effectiveness. It typically performs in the top tier for speed (71st percentile) and offers competitive pricing (74th percentile). The model exhibits strong reliability with a 95% success rate across benchmarks, indicating consistent operational stability. In terms of specific capabilities, Hermes 4 70B excels in hallucination prevention (98.0% accuracy) and email classification (98.0% accuracy), showcasing its ability to handle uncertainty and categorize information effectively. Its performance in ethics (99.0% accuracy) is also very strong. While general knowledge is solid (96.5%), its accuracy in instruction following (55.1%), coding (55.0%), reasoning (54.0%), and mathematics (69.0%) suggests areas for potential improvement, particularly given its emphasis on verified reasoning data. The model's hybrid reasoning mode and support for structured outputs like JSON and function calling are key features designed to enhance steerability and reduce refusal rates.