Atlas

Models

← All models

Hermes 4 405B

Nous Research · 2025-08-26 · 405.9B parameters

Hermes 4 405B is Nous Research's hybrid-mode reasoning model, a full-parameter fine-tune of Llama-3.1-405B released in August 2025. It introduces toggleable extended thinking via <think> tags and was trained on roughly 5 million samples and 60 billion post-training tokens spanning verified reasoning traces, math, code, and general instruction data. Because those corpus figures are approximate, no exact training-token count is recorded.

Benchmark scores

BenchmarkScore
AA-LCR20.7
AA-LCR v1.122.3
AA-Omniscience Index-33.3
AIME 2025 (Artificial Analysis avg@10)69.7
Artificial Analysis Omniscience Accuracy30.1
Artificial Analysis Omniscience Hallucination Rate80.2
Artificial Analysis Openness Index47.2
Artificial Analysis Progress Index v1207
BuseyBench SVG2.0
Capability133.9
Creative Writing v3 Elo1421
Creative Writing v3 Rubric Score16.1
CritPt (avg@5)0.3
EQ-Bench 31361
EQ-Bench 3 (Rubric Score)65.2
EQ-Bench 3: Analytic8.6
EQ-Bench 3: Assertive6.2
EQ-Bench 3: Compliant5.8
EQ-Bench 3: Empathy7.1
EQ-Bench 3: Humanlike5.7
EQ-Bench 3: Insight7.1
EQ-Bench 3: Moralising4.8
EQ-Bench 3: Pragmatic7.0
EQ-Bench 3: Safety5.7
EQ-Bench 3: Social IQ6.5
EQ-Bench 3: Warm6.6
GPQA Diamond (Artificial Analysis avg@5)72.7
Humanity's Last Exam (Text-Only)10.9
IFBench (Artificial Analysis avg@5)34.8
LiveCodeBench (Artificial Analysis)68.6
Loading Atlas data…