Hermes 4 is a large-scale reasoning model built on Meta-Llama-3.1-405B and released by Nous Research. It introduces a hybrid reasoning mode, where the model can choose to deliberate internally with <think>...</think> traces or respond directly, offering flexibility between speed and depth. The model is instruction-tuned with an expanded post-training corpus (~60B tokens) emphasizing reasoning traces, improving performance in math, code, STEM, and logical reasoning, while retaining broad assistant utility. It also supports structured outputs, including JSON mode, schema adherence, function calling, and tool use. Hermes 4 is trained for steerability, lower refusal rates, and alignment toward neutral, user-directed behavior.
Evaluations
13
across 12 benchmarks
Latency
461ms
Context length
131k
tokens
Cost
$0.30 · $1.20
input · output per 1M tokens
Key takeaways
Hermes 4 is a large-scale, Transformer-based reasoning model for instruction following, creative generation, question answering, and tool use, featuring a hybrid reasoning mode for balancing speed and depth.
It was instruction-tuned with an expanded post-training corpus (~60B tokens) emphasizing reasoning traces to improve performance in math, code, STEM, and logical reasoning, and supports structured outputs including JSON and function calling.
The model demonstrates strong general performance in instruction following, creative writing, and maintaining low toxicity (75.4% accuracy on 'IFEval - General LLM Abilities Benchmark') and excels in subjects requiring direct factual recall, particularly in medical and scientific fields (77.7% accuracy on MMLU).
Hermes 4 struggles significantly with complex multi-step reasoning, precise numerical calculations, and strict adherence to intricate formatting guidelines, as evidenced by its 3.3% accuracy on 'AIME 2025' for structured scientific reasoning.
Common failure modes include misinterpreting problem constraints, incorrect mathematical calculations, logical errors in multi-step deductions, and failing to follow negative constraints or exact output formatting, particularly in complex ethical reasoning and highly specialized technical subjects like abstract algebra.
