Article contents:
- Introduction: a revolution in open artificial intelligence
- “Hybrid reasoning” logic: how does Hermes 4 excel?
- Behind the scenes: revolutionary training systems
- Nous Research philosophy: why safety restrictions are “annoying”
- How does a startup compete with tech giants?
- The technical fix for the endless thinking problem
- How to try Hermes 4 and its cost
- Conclusion: the future of artificial intelligence between freedom and responsibility
Introduction: a revolution in open artificial intelligence
Nous Research, an enigmatic AI startup that has emerged as a leading voice in the open-source AI movement, quietly launched “Hermes 4” on Monday, a collection of large language models that the company claims can match leading proprietary systems while offering unprecedented user control and minimal content restrictions. This release marks a major escalation in the battle between open-source AI advocates and big tech companies over who should control access to advanced AI capabilities. Unlike models from OpenAI, Google, or Anthropic, Hermes 4 is designed to respond to almost any request without the safety guards that have become standard in commercial AI systems.
“Hybrid reasoning” logic: how does Hermes 4 excel?
Hermes 4 introduces what Nous Research calls “hybrid reasoning,” allowing users to switch between fast responses and deeper step-by-step thinking processes. When activated, the models generate their internal logic within special tags (<think>) before delivering a final answer—similar to OpenAI’s o1 reasoning models but with full transparency into the AI’s thought process. In testing, the largest Hermes 4 model with 405 billion parameters scored 96.3% on the MATH-500 benchmark in reasoning mode and 81.9% on the difficult AIME’24 math competition—performance that rivals or exceeds many proprietary systems that cost millions to develop. Crucially, Hermes 4 achieved the highest score among all tested models on “RefusalBench,” a new benchmark created by Nous Research to measure how often AI systems refuse to answer questions, scoring 57.1% and significantly outperforming GPT-4o (17.67%).
Behind the scenes: revolutionary training systems
Behind the capabilities of Hermes 4 lies a sophisticated training infrastructure. The models were trained using two new systems: DataForge, a graph-based synthetic data generator, and Atropos, an open-source reinforcement learning framework. DataForge generates training data through what the company describes as “random walks” across directed graphs, transforming simple training data into complex instruction-following examples. Meanwhile, Atropos operates like hundreds of specialized training environments where AI models practice specific skills such as math, coding, tool use, and creative writing, receiving feedback only when they produce correct solutions.
Nous Research philosophy: why safety restrictions are “annoying”
Nous Research has built its reputation on a philosophy that puts user control above corporate content policies. Its models are designed to be “steerable,” meaning they can be modified or guided to behave in specific ways without the strict safety constraints that characterize commercial AI systems. Tommy Shaughnessy, a venture capitalist at Delphi Ventures who invested in Nous Research, wrote: “Hermes 4 is not burdened by the disclaimers, rules, and hyper-caution that is so annoying and harms innovation and usability.” This approach puts the company at the center of the ongoing debate over AI safety, with the firm arguing that transparency and user control are better than institutional guardianship.
How does a startup compete with tech giants?
The launch of Hermes 4 comes at a pivotal moment in the AI industry. While big tech companies have invested billions, a growing open-source movement shows that these capabilities should not be controlled by a handful of corporations. Nous Research, which raised $65 million earlier this year, demonstrates that specialized techniques can compete with the massive scale of tech giants. Training the largest model required 192 Nvidia B200 GPUs and 71,616 GPU hours—a significant computational investment, but not unprecedented.
The technical fix for the endless thinking problem
One of Hermes 4’s most important technical contributions is addressing a problem plaguing reasoning models: excessively long thought processes. Researchers found that their smaller 14-billion-parameter model reaches the maximum context length 60% of the time when reasoning, meaning it gets stuck in endless thinking loops. Their solution involved a second training phase that teaches models to stop thinking at exactly 30,000 tokens, reducing long generation by 65–79% while preserving most reasoning performance.
Conclusion: the future of artificial intelligence between freedom and responsibility
The launch of Hermes 4 represents more than just another AI model release—it is a statement about who should control the future of artificial intelligence. The company’s approach raises fundamental questions about the tradeoff between safety and capability, and between corporate control and user freedom. Hermes 4 has proven that the future of AI will not be determined solely by companies with the largest budgets.
