Nous Research releases Hermes 4

Nous Research released Hermes 4, a family of open-weight language models at 14B, 70B and 405B parameters, built by post-training Meta’s Llama 3.1 checkpoints rather than pretraining from scratch. The models offered a hybrid reasoning mode, toggling between a fast standard response and an explicit step-by-step reasoning process wrapped in visible <think> tags for harder problems. Nous reported strong benchmark results for the 405B model, including scores in the high 70s and 80s percent on AIME-style maths competition problems and GPQA Diamond, achieved through post-training refinement of an existing base model rather than new pretraining compute.

The release’s distinguishing claim was around refusal behaviour: Nous reported the 405B model scoring 57.1% on RefusalBench, a benchmark measuring willingness to answer legitimate but sensitive queries, against 17.67% for GPT-4o and a similarly low score for Claude Sonnet 4. Nous framed this as evidence of a deliberately “neutral alignment” approach, aiming for a model steerable by the user rather than one with the more restrictive default guardrails built into major labs’ consumer products, while still declining clearly harmful requests.

Weights and a lengthy technical report were published openly, continuing Nous Research’s position as one of the most active publishers of frontier-adjacent open-weight models built primarily through post-training technique rather than large-scale pretraining. Hermes 4’s approach — and its explicit contrast with the refusal rates of closed frontier models — kept alive a running argument in the open-weight community over whether low refusal rates represented meaningful user autonomy or reduced safety margin, an argument that predated Hermes 4 and continued through Nous’s subsequent releases.