Nous Research Hermes 4 AI Challenges ChatGPT Without Restrictions
In a significant move for the open-source AI community, the secretive startup Nous Research has released Hermes 4, a new family of large language models. The company claims these models can rival the performance of top proprietary systems while offering users greater control and virtually no content restrictions.
This release intensifies the ongoing debate between open-source AI proponents and major tech corporations over the control of advanced AI. Unlike models from industry giants like OpenAI, Google, or Anthropic, Hermes 4 is intentionally designed to bypass the standard safety guardrails found in commercial AI, responding to a much wider range of prompts.
In their announcement on X (formerly Twitter), Nous Research stated, “Special attention was given to making the models creative and interesting to interact with, unencumbered by censorship, and neutrally aligned while maintaining state of the art level math, coding, and reasoning performance for open weight models.”
Hermes 4's Hybrid Reasoning Outperforms ChatGPT
Hermes 4 introduces a feature called “hybrid reasoning,” which allows users to switch between quick responses and a more detailed, step-by-step thinking process. When this mode is active, the model reveals its internal reasoning within special <think> tags before delivering a final answer. This provides a transparent look into the AI's thought process, similar in concept to OpenAI’s o1 reasoning models but fully visible to the user.
The model's performance on technical benchmarks is impressive. The largest 405-billion parameter version of Hermes 4 scored 96.3% on the MATH-500 benchmark and 81.9% on the difficult AIME’24 mathematics competition when in reasoning mode. These scores are competitive with, or even surpass, many proprietary systems that cost significantly more to develop.
Critically, Hermes 4 achieved the top score on “RefusalBench,” a new test designed by Nous Research to measure how frequently an AI model refuses to answer a prompt. The model scored 57.1% in reasoning mode, far outperforming GPT-4o (17.67%) and Claude Sonnet 4 (17%).
The Tech Behind the Triumph: DataForge and Atropos
The power of Hermes 4 is rooted in a sophisticated training infrastructure developed by Nous Research. The models were trained using two innovative systems: DataForge, a synthetic data generator, and Atropos, an open-source reinforcement learning framework.
DataForge generates complex instruction-following examples from simple data. For instance, it can convert a Wikipedia article into a rap song and then create questions and answers based on the new content. Atropos functions like a series of specialized training gyms where the AI models practice skills like math and coding, receiving feedback only on correct solutions. This “rejection sampling” method ensures that the final training dataset consists of only high-quality, verified responses.
Tommy Shaughnessy, a venture capitalist at Delphi Ventures and an investor in Nous Research, explained, “All in the dataset contains 3.5 million reasoning samples and 1.6 million non-reasoning samples! Hermes was trained on RL data, not just static datasets of question and answer!”
A Philosophy of Freedom: Uncensored by Design
Nous Research has built its brand on prioritizing user control over corporate content policies. Their models are designed to be “steerable,” allowing users to fine-tune them for specific behaviors without the rigid safety filters common in commercial AI.
“Hermes 4 is not shackled by disclaimers, rules and being overly cautious which is annoying as hell and hurts innovation and usability,” wrote Shaughnessy. “If its open source but refuses all requests its pointless. Not an issue with Hermes 4.”
This uncensored approach has made Nous Research popular with developers seeking maximum flexibility, but it also places the company in the middle of debates on AI safety. While acknowledging the potential for misuse, Nous Research advocates for transparency and user control as a better alternative to corporate gatekeeping.
Startup vs. Goliath: How Nous Research Competes with Big Tech
Hermes 4 arrives as the open-source AI movement is gaining serious momentum. Models like Meta’s Llama 3.1 and others from DeepSeek and Alibaba are increasingly competitive with proprietary systems. Hermes 4 pushes this trend further, especially in the realm of complex reasoning.
As Shaughnessy noted, “Nous is a startup with dozens of extremely talented people... they continue to put out innovative models and research at an insane pace.” The startup, which recently raised $65 million, is proving that specialized techniques can compete with the massive budgets of tech giants.
Solving the Endless Loop: A Key Technical Breakthrough
One of the most significant technical achievements of Hermes 4 is a solution to a common problem in reasoning models: getting stuck in endless thinking loops. The researchers discovered that their smaller models would often exhaust their context limits while reasoning. To fix this, they implemented a second training stage that taught the models to stop reasoning at a specific token count, which reduced overlong generations by up to 79% without sacrificing much performance. This “length control” technique is a valuable contribution to the broader AI research community.
How to Access Hermes 4
In line with its open-source ethos, Nous Research has made Hermes 4 widely available. The model weights can be downloaded from Hugging Face, and API access is available through the company's new chat interface and various inference providers. This makes the models an attractive alternative for enterprises and researchers who need customization or wish to avoid the costs and restrictions of proprietary APIs.
The Bigger Picture: Reshaping the Future of AI
The launch of Hermes 4 is more than a technical achievement; it's a statement on the future of AI control. Nous Research has shown that a small, innovative team can challenge the dominance of tech giants. Their work forces a conversation about the balance between safety and capability, and between corporate oversight and user freedom.
While the long-term impact of this philosophy is yet to be determined, Hermes 4 proves that the future of AI will not be dictated solely by the companies with the largest budgets.
