Loading the catalogue…
Loading the catalogue…
NousResearch is a private, US-incorporated open-weights lab that publishes Hermes model weights freely under mixed licences (Llama Community, MIT, Seed, Qwen terms), with a deliberate design philosophy of reduced refusals and high steerability that drives strong open-source adoption but creates EU AI Act transparency gaps. It is fully CLOUD Act and FISA §702 exposed with no EU legal entity and no demonstrated GDPR compliance posture, and its hosted Nous Portal privacy policy has previously stated that user conversations are collected for model training. Self-hosting Hermes weights in EEA infrastructure avoids most of this data-path risk, but EU regulated-sector deployers must add their own content-governance layer given the models' minimal built-in guardrails.
US-incorporated entity fully exposed to CLOUD Act compulsion and likely FISA §702 orders. No EU legal entity or EU data-processing infrastructure identified, and no SCCs or adequacy-based transfer mechanism published.
No EU AI Act compliance statement, GPAI Code of Practice signature, or structured Article 53 transparency template found, despite GPAI obligations having been enforceable since 2 August 2025.
The Nous Portal privacy policy has previously stated that user conversation data is collected and used to train models, with no GDPR lawful basis, DPO, or opt-out mechanism described. EU enterprises using the hosted portal (as opposed to self-hosted weights) face direct data-transfer risk.
Hermes models are deliberately designed with minimal content guardrails, scoring far higher than mainstream aligned models on the RefusalBench benchmark. EU regulated-sector deployers must add their own content-governance and output-filtering layer.
No SOC 2, ISO 27001, or equivalent security certification, and no formal bug bounty or responsible disclosure programme found published anywhere on the official site.
Hermes 4 fine-tunes are subject to the Meta Llama Community Licence (restricting training of competing foundation models and imposing a commercial MAU threshold); Hermes 4.3 and NousCoder-14B carry separate ByteDance Seed and Qwen licence terms respectively, creating a fragmented licensing landscape for redistribution or fine-tuning.
Stav’s assessment
Editorial assessment, not legal advice. Stav's risk ratings, scores, and verdicts are our own analysis of publicly available information and may be incomplete or out of date. Verify independently before making compliance or procurement decisions.
Published a 94-page Hermes 4 Technical Report (arXiv:2508.18255) with detailed training-data composition, benchmark methodology, and evaluation data, exceeding typical open-source lab disclosure norms.
All Hermes model weights are released publicly on HuggingFace, enabling independent evaluation and fully self-hosted EU deployment with no data path to NousResearch's own servers.
Hermes Agent is reported to have crossed 140,000 GitHub stars within about five months of launch and was described as the most-used agent on OpenRouter by mid-2026, indicating strong independent community scrutiny.
Sustained, high-velocity release cadence through July 2026 (v0.18.0 clearing all outstanding P0/P1 issues), NVIDIA GTC keynote featuring Hermes Desktop, and integration with NVIDIA RTX/DGX Spark hardware — no indications of contraction or wind-down.
Published safeguards & certifications