Paper 2508.18255
Hermes 4 Technical Report
- Published
- Aug 2025
- Research lab
- Independent
- Citations
- 3
- GitHub
- Not linked
01 In brief
Summary
Hermes 4 is a family of hybrid reasoning models (14B, 70B, 405B) that combine structured, multi-turn reasoning with broad instruction-following.
The models were trained on a dataset of ~5 million samples (19B tokens), including 3.5M reasoning and 1.6M non-reasoning samples, synthesized via DataForge, a graph-based generator, and rejection sampling with task-specific verifiers.
Training used loss-masking, length-control fine-tuning, and efficient packing, with a second SFT stage to cap reasoning at 30k tokens, reducing overlong rates by over 98% with minimal benchmark impact.
Evaluations across math, code, knowledge, and alignment show Hermes 4 405B achieving strong results (e.g., AIME'24 81.9, GPQA Diamond 70.6, LCBv6 61.4) and high RefusalBench scores (57.1), indicating fewer refusals.
Qualitative analysis shows greater behavioral plasticity, including better persona adoption and reduced sycophancy under prompt modifications.
All weights are publicly released.
02 From the paper
Abstract
We present Hermes 4, a family of hybrid reasoning models that combine structured, multi-turn reasoning with broad instruction-following ability. We describe the challenges encountered during data curation, synthesis, training, and evaluation, and outline the solutions employed to address these challenges at scale. We comprehensively evaluate across mathematical reasoning, coding, knowledge, comprehension, and alignment benchmarks, and we report both quantitative performance and qualitative behavioral analysis. To support open research, all model weights are published publicly at https://huggingface.co/collections/NousResearch/hermes-4-collection-68a731bfd452e20816725728