Nous Research falls Hermes 4 AI models that surpass Chatgpt without content restrictions

“The challenge is making the effects of thinking use Rohan Paul on xHighlighting one of the technical breakthroughs in the version.

Perhaps the most prominent, Hermes 4 He achieved the highest degree of all models tested on “Refusalbench”, a new research created to measure the number of times that artificial intelligence systems refuse to answer questions. The model recorded 57.1 % in thinking mode, greatly outperforms GPT-4O (17.67 %) and Clauds Sonnet 4 (17 %).

Inside DatAforge and Atropos: Training systems penetration behind Hermes 4 capabilities

Behind the capabilities of Hermes 4 lies an advanced training infrastructure Nous search It has evolved over several years. Models have been trained using two new systems: DataforgeIndustrial data generator based on the graph, and AtroposFire of reinforcement open source.

Dataforge It creates training data through what the company describes as “random aspects” through guided graphics, which transforms simple training data into follow -up examples of complex instructions. The system, for example, can take the Wikipedia article and turn it into a rap song, then create questions and answers based on this transformation.

AtroposMeanwhile, it works like hundreds of specialized training environments where artificial intelligence models practice specific skills – Mathematics, coding, use of tools, and creative writing – comments only when producing correct solutions. The “Rejection samples” approach ensures that only high -quality verified responses make them in training data.

ATROPOS is the reinforcement learning frame

ATROPOS is an environment for learning to enhance open sources by NOUS which contains hundreds of “gyms” (such as mathematics, coding, games, tools, and vision) to train and evaluate LLM paths via RL sessions.

In other words … pic.twitter.com/fjxaqklez

– Tommy (@shaghnessy119) August 26, 2025

“Nous these environments to create a data collection for Hermes 4!” Make up Tommy ShunisiProject capitalism in Delphi Ventures Which invested in Nous Research. “Everything in the data set contains 3.5 million thinking samples and 1.6 million non -widespread samples! Hermes have been trained on RL data, not just fixed data sets of questions and answer!”

The training process requires 192 Nvidia B200 graphics processing units And 71,616 GPU hours of the largest model – an important account investment but unprecedentedly explains how specialized technologies can compete with the huge range of technology giants.

Why Nous Research believes that handrails of artificial intelligence are “annoying like hell” and hurt innovation

Nous search She built her reputation on a philosophy that places the user control over the company’s content policies. The company’s models are designed to be “directed”, which means that they can be caught or demanded to act in specific ways without solid safety restrictions that characterize commercial artificial intelligence systems.

“Hermes 4 is not overcome before the evacuation of responsibility, rules, excessive caution that bothering hell and hurts innovation and ease of use,” said Shunisi in a detailed thread. “If the source is open, but it rejects all the meaningless requests. It is not a problem with Hermes 4.”

Hermes 4 has not been killed due to the evacuation of responsibility, rules, and excessive cautious caution that disturbs hell, hurts innovation and ease of use.

Hermes 4 70B is unlike the Spectrum Vs Openai Open Source. It is also ~ 4x more open versus ChatGPT 4O!

If it is open … pic.twitter.com/q5rpx1oozo

– Tommy (@shaghnessy119) August 26, 2025

He took this approach Nous search It is common among researchers and developers who want the maximum flexibility, but also puts the company at the current discussion center on the integrity of artificial intelligence and moderate content. While models can theoretically use for harmful purposes, Nous Research argues that transparency and controlling user are preferred to maintain corporate gate.

Company Technical report,, Which was released alongside the models, provides unprecedented details about the training process, evaluation results, and even the actual text outputs of the standard tests. “We believe that this report defines a new standard for transparency in measurement,” the company said.

How a small startup competes with 192 graphics processing units with Big Tech Budgets of AI is billions of dollars

Hermes 4A version comes in a pivotal moment in the artificial intelligence industry. While major technology companies have poured billions of dollars to develop increasingly increasingly artificial intelligence systems, an increasing open source movement argues that these capabilities should not be controlled by a handful of companies.

The last months have witnessed great developments in open source artificial intelligence, with models such as Meta’s Lama 3.1and Deepseek’s R1And Ali Baba Qwen The performance achievement series that competes with property systems. Hermes 4 represents another step in this progress, especially in the field of thinking – as it is long power for closed systems Openai’s O1.

“First, Nous is a startup that includes dozens of people very talented”, ” Note Shunisi. “They do not have an annual Capex spending worth $ 100 billion to buy Easter or 1,000 employees, and although they continue to launch innovative models and crazy pace.”

Startup, which Rapid $ 65 million in financing Earlier this year led modelAnd it was developed PsycheA distributed training system aimed at coordinating artificial intelligence training via Internet computers using Blockchain technology.

Technical reform that prevented Hermes 4 from thinking about endless episodes

One Hermes 4The most important technical contributions addresses the thinking models: excessive long thinking processes. The researchers have found that the microcredit model of 14 billion people will reach the maximum period of 60 % of the time when thinking is, and they mainly stop in endless episodes of thinking.

Their solution guarantees a second training phase that teaches models to stop thinking about exactly 30,000 symbols, which reduces the cruise generation by 65-79 % while maintaining most of the thinking performance. The “length control” technique can be valuable for the broader artificial intelligence research community.

“Smaller models (<14B) tend to think about thinking when distilled, but larger models do Muyu is on xShedding light on visions of the technical report.

but, Hermes 4 She still faces common restrictions in open source models. Despite impressive standard performance, models require large mathematical resources to run them and may not coincide with ease of use or reliability of artificial intelligence services for many applications.

Where do Hermes 4 try and what it costs compared to Chatgpt and Claud

Nous search I have made Hermes 4 Available through multiple channels, which reflects the open source philosophy. Typical weights can be downloaded freely in the face of the embrace, while the company also provides access to the application programming interface through the renovated chat interface and partnerships with inference providers such as Chues, Nebius and Luminal.

“You can try Hermes 4 in the new renewable Nous Nous user interface,” the company announced, highlighting features such as parallel reactions and memory system.

For the institution’s users and researchers, the models represent an attractive, potential alternative to pay the API access to property systems, especially for applications that require high levels of customization or dealing with sensitive content.

The biggest picture: What does Hermes 4 mean for the future of artificial intelligence development

release Hermes 4 More than just launching another AI model – it is a statement about who should control the future of artificial intelligence. In an increasingly dominated industry a handful of technology giants with almost unlimited resources, Nous Research has proven that innovation still could come from unexpected places.

The company’s approach raises basic questions about the differentials between safety and ability, between companies control and user freedom. While major technology companies argue that moderate content and accurate safety studies are necessary to spread responsible artificial intelligence, Nous Research claims that transparency and the user agency are more important than the restrictions imposed on companies.

Whether this philosophy will eventually prove useful or problematic that we should see. But there is one certain thing: Hermes 4 has shown that the future of artificial intelligence will not be determined only by companies that have the deepest pockets.

In a field where yesterday’s impossible goods become goods, Nous Research has proven that the only most dangerous thing of artificial intelligence that says cannot be ready to say yes.

Daily visions about business use cases with VB daily

If you want to persuade your boss at work, you have covered VB Daily. We give you the internal journalistic precedence over what companies do with obstetric artificial intelligence, from organizational transformations to practical publishing operations, so that you can share visions of the maximum return on investment.

Read with us privacy policy

Thanks for subscribing. Check more VB news bulletins here.

An error occurred.