Nous Analysis drops Hermes 4 AI fashions that outperform ChatGPT with out content material restrictions

“The problem is making pondering traces helpful and verifiable with out runaway reasoning,” famous AI researcher Rohan Paul on X, highlighting one of many technical breakthroughs within the launch.

Maybe most notably, Hermes 4 achieved the very best rating amongst all examined fashions on “RefusalBench,” a brand new benchmark Nous Analysis created to measure how typically AI programs refuse to reply questions. The mannequin scored 57.1% in reasoning mode, considerably outperforming GPT-4o (17.67%) and Claude Sonnet 4 (17%).

Inside DataForge and Atropos: The breakthrough coaching programs behind Hermes 4’s capabilities

Behind Hermes 4’s capabilities lies a classy coaching infrastructure that Nous Analysis has developed over a number of years. The fashions have been skilled utilizing two novel programs: DataForge, a graph-based artificial information generator, and Atropos, an open-source reinforcement studying framework.

DataForge creates coaching information by what the corporate describes as “random walks” by directed graphs, remodeling easy pre-training information into complicated instruction-following examples. The system can, for example, take a Wikipedia article and rework it right into a rap music, then generate questions and solutions based mostly on that transformation.

Atropos, in the meantime, operates like a whole bunch of specialised coaching environments the place AI fashions observe particular abilities—arithmetic, coding, device use, and artistic writing—receiving suggestions solely after they produce appropriate options. This “rejection sampling” strategy ensures that solely verified, high-quality responses make it into the coaching information.

Atropos is Nous’ Reinforcement Studying framework

Atropos is an open supply reinforcement studying setting by Nous that has a whole bunch of “gyms” (like math, coding, video games, device‑use, imaginative and prescient) to coach and consider LLM trajectories through scalable, async RL loops.

In different phrases… pic.twitter.com/fjxaQKClEZ

— Tommy (@Shaughnessy119) August 26, 2025

“Nous used these environments to generate the dataset for Hermes 4!” defined Tommy Shaughnessy, a enterprise capitalist at Delphi Ventures who has invested in Nous Analysis. “All within the dataset comprises 3.5 million reasoning samples and 1.6 million non-reasoning samples! Hermes was skilled on RL information, not simply static datasets of query and reply!”

The coaching course of required 192 Nvidia B200 GPUs and 71,616 GPU hours for the most important mannequin — a major however not unprecedented computational funding that demonstrates how specialised methods can compete with the large scale of tech giants.

Why Nous Analysis believes AI security guardrails are ‘annoying as hell’ and harm innovation

Nous Analysis has constructed its popularity on a philosophy that places person management above company content material insurance policies. The corporate’s fashions are designed to be “steerable,” which means they are often fine-tuned or prompted to behave in particular methods with out the inflexible security constraints that characterize business AI programs.

“Hermes 4 is just not shackled by disclaimers, guidelines and being overly cautious which is annoying as hell and hurts innovation and value,” wrote Shaughnessy in an in depth thread analyzing the discharge. “If its open supply however refuses all requests its pointless. Not a problem with Hermes 4.”

Hermes 4 is just not shackled by disclaimers, guidelines and being overly cautious which is annoying as hell and hurts innovation and value.

Hermes 4 70B is at the exact opposite of the spectrum vs OpenAI’s open supply mannequin. It is also ~4x extra open vs ChatGPT 4o!

If its open… pic.twitter.com/q5RpX1oOzo

— Tommy (@Shaughnessy119) August 26, 2025

This strategy has made Nous Analysis standard amongst AI researchers and builders who need most flexibility, nevertheless it additionally locations the corporate on the middle of ongoing debates about AI security and content material moderation. Whereas the fashions can theoretically be used for dangerous functions, Nous Analysis argues that transparency and person management are preferable to company gatekeeping.

The corporate’s technical report, launched alongside the fashions, supplies unprecedented element concerning the coaching course of, analysis outcomes, and even the precise textual content outputs from benchmark assessments. “We consider this report units a brand new customary for transparency in benchmarking,” the corporate acknowledged.

How a small startup with 192 GPUs is competing towards Massive Tech’s billion-dollar AI budgets

Hermes 4‘s launch comes at a pivotal second within the AI business. Whereas main know-how firms have poured billions into growing more and more highly effective AI programs, a rising open-source motion argues that these capabilities shouldn’t be managed by a handful of firms.

Current months have seen vital advances in open-source AI, with fashions like Meta’s Llama 3.1, DeepSeek’s R1, and Alibaba’s Qwen sequence attaining efficiency that rivals proprietary programs. Hermes 4 represents one other step on this development, significantly within the space of reasoning—lengthy thought of a power of closed programs like OpenAI’s o1.

“First up, Nous is a startup with dozens of extraordinarily proficient individuals,” famous Shaughnessy. “They don’t have the $100b+ annual capex spend of a hyperscaler nor 1,000’s of workers and regardless of that they proceed to place out modern fashions and analysis at an insane tempo.”

The startup, which raised $65 million in funding earlier this yr led by Paradigm, has additionally been growing Psyche Community, a distributed coaching system that goals to coordinate AI coaching throughout internet-connected computer systems utilizing blockchain know-how.

The technical repair that stopped Hermes 4 from pondering in infinite loops

One in every of Hermes 4‘s most vital technical contributions addresses an issue plaguing reasoning fashions: overly lengthy pondering processes. The researchers discovered that their smaller 14-billion parameter mannequin would attain most context size 60% of the time when reasoning, primarily getting caught in infinite loops of pondering.

Their answer concerned a second coaching stage that teaches fashions to cease reasoning at precisely 30,000 tokens, lowering overlong technology by 65-79% whereas sustaining a lot of the reasoning efficiency. This “size management” method may show useful for the broader AI analysis group.

“Smaller fashions (<14B) are likely to overthink when distilled, however bigger fashions don’t,” noticed AI researcher Muyu He on X, highlighting insights from the technical report.

Nonetheless, Hermes 4 nonetheless faces limitations frequent to open-source fashions. Regardless of spectacular benchmark efficiency, the fashions require vital computational assets to run and will not match the convenience of use or reliability of economic AI companies for a lot of purposes.

The place to attempt Hermes 4 and what it prices in comparison with ChatGPT and Claude

Nous Analysis has made Hermes 4 obtainable by a number of channels, reflecting the open-source philosophy. The mannequin weights are freely downloadable on Hugging Face, whereas the corporate additionally gives API entry by its revamped chat interface and partnerships with inference suppliers like Chutes, Nebius, and Luminal.

“You possibly can attempt Hermes 4 within the new, revamped Nous Chat UI,” the corporate introduced, highlighting options like parallel interactions and a reminiscence system.

For enterprise customers and researchers, the fashions characterize a probably enticing various to paying for API entry to proprietary programs, particularly for purposes requiring excessive ranges of customization or dealing with of delicate content material.

The larger image: What Hermes 4 means for the way forward for AI improvement

The discharge of Hermes 4 represents extra than simply one other AI mannequin launch — it’s a press release about who ought to management the way forward for synthetic intelligence. In an business more and more dominated by a handful of tech giants with nearly limitless assets, Nous Analysis has demonstrated that innovation can nonetheless come from surprising locations.

The corporate’s strategy raises elementary questions concerning the trade-offs between security and functionality, between company management and person freedom. Whereas main know-how firms argue that cautious content material moderation and security guardrails are important for accountable AI deployment, Nous Analysis contends that transparency and person company are extra vital than corporate-imposed restrictions.

Whether or not this philosophy will in the end show useful or problematic stays to be seen. However one factor is definite: Hermes 4 has proven that the way forward for AI gained’t be decided solely by the businesses with the deepest pockets.

In a area the place yesterday’s impossibilities grow to be tomorrow’s commodities, Nous Analysis simply proved that the one factor extra harmful than an AI that claims no is likely to be one which’s prepared to say sure.

Each day insights on enterprise use circumstances with VB Each day

If you wish to impress your boss, VB Each day has you coated. We provide the inside scoop on what firms are doing with generative AI, from regulatory shifts to sensible deployments, so you’ll be able to share insights for max ROI.

Learn our Privateness Coverage

Thanks for subscribing. Try extra VB newsletters right here.

An error occured.