Nous Research left the Hermes 4AI model that performs the chat without material restrictions.

The AI ​​researcher said, “The challenge is telling Rohan Paul X on XTo highlight one of the technical successes in the release.

Perhaps most especially, Hemes 4 Gained the highest score among all tested models on “Refuselvench”, a new benchmark seal research that was designed to measure how many times the AI ​​systems refuse to answer questions. The model scored 57.1%in Reasoning Mode, making GPT-4o (17.67%) and Cloud Sonnet 4 (17%) better.

Inside Dataforge and Atropos: Success Training System behind the capabilities of Hermes 4

Hermes 4 capabilities are a sophisticated training infrastructure Sewage research Has developed over many years. The model was trained using two novel systems: DataforgeA graph-based synthetic data generator, and AtolosThe outline of learning an open-source reinforcement.

Dataforge Through the “random walk” through the graph directed by the company, the training data is made through it, the simple pre-training data transforms data into complex instructions-the following examples. For example, the system can take a Wikipedia article and turn it into a rap song, then generates questions and answers based on that change.

AtolosMeanwhile, hundreds of special training operates like environment, where AI models practice specific skills – analogy, coding, use of tools, and creative writing – only they receive response only when they produce the correct solution. This “rejection sampling” approach ensures that only verified, high-quality reactions make it in training data.

Atropos nous’ reinforcement is outline

An open source reinforcement by Atropos Nous is a learning environment that has hundreds of “gym” (eg mathematics, coding, games, tools, uses, vision), which are to train and evaluate LLM trajectory through scalable, AsyNC RL loops.

In other words… pic.twitter.com/fjxaqkclez

– Tommy (@Shaughnessy119) 26 August, 2025

“Nose used these environments to generate datasets for Hermes 4!” Explained Tommy ShaogheniOn an enterprise capitalist Deli ventures Those who have invested in sewage research. “All in the dataset have 3.5 million argument samples and 1.6 million non-purpose specimens! Hermes were trained on RL data, not just questions and stable datasets of answers!”

The training process requires 192 nvidia B200 gpus And 71,616 GPU hours for the largest model – not an important but unprecedented computational investment that shows how special technology can compete extensively of technical giants.

Why Nous Research believes that AI security guards are ‘annoying as hell’ and hurt innovation

Sewage research A philosophy has created its reputation that holds user control over corporate material policies. The company’s model is designed to be “steerable”, which means they may be fine or motivated to behave in specific ways without rigorous safety obstacles that characterize the commercial AI system.

“Hermes 4 is not done by disgrace, rules and excessive being alert that is annoying as hell and damages innovation and purpose,” toilet has written in a wide thread analyzing the release. “If it has an open source, but refuses all requests. Hermes is not an issue with 4.”.

Hermis 4 is not brought down from disgrace, rules and excessive being alert that is annoying as hell and damages innovation and purpose.

Hermes 4 70B spectrum vs. Openai’s Open Source model is the complete opposite. It is also ~ 4x more open vs. chat 4o!

If it open … pic.twitter.com/q5rpx1oozo

– Tommy (@Shaughnessy119) 26 August, 2025

This approach has created Sewage research The AI ​​who wants maximum flexibility among researchers and developers, but also keeps the company at the center of the ongoing debate about AI security and material moderation. While the model can be theoretically used for harmful purposes, no rational research argues that transparency and user control are better for corporate gatekeeping.

Company technical ReportThe model provides unprecedented details about the release, training process, evaluation results and even the actual text output from benchmark tests. “We believe this report sets a new standard for transparency in benchmarking,” the company said.

A small startup with 192 GPU is competing against Big Tech’s Billion-Dollar AI budget

Hemes 4The release release in a significant moment in the AI ​​industry. While major technology companies have poured billions in developing powerful AI systems rapidly, a growing open-source movement argues that these capabilities should not be controlled by a handful of corporations.

In recent months, significant progress has been observed in Open-SOS AI, which have models like meta Lama 3.1, Deepsek’s R1And Alibaba’s Caven The series receiving the series is rival to rival proprietary systems. Hermis 4 represents another step in this progress, especially in the field of logic – such as long -closed system strength is considered OPENAI’s O1,

“First of all, Nous is a startup with dozens of highly talented people,” Noted toilet“They do not have a $ 100B+ annual Capex of Hypersscaleer and

Startup, which Gather $ 65 million in funding Leaded earlier this year IdealIs also developing Manas NetworkA distributed training system aims to coordinate AI training in computers connected to the Internet using blockchain technology.

Technical fix that stopped Hermes 4 from thinking in endless loops

One of one Hemes 4The most important technical contribution addresses a problem due to a problem: prolonged thinking processes. Researchers found that their small 14-Billion parameter model would reach maximum reference length 60% time, when logic, essentially trapped in endless loops of thinking.

Their solution included a second training phase that teaches the model to prevent arguments on a fine 30,000 tokens, which reduces the overload generation by 65–79% while maintaining most of the performance of logic of logic. This “length control” technology can prove to be valuable for the broader AI research community.

“Small models (<14b) overthrow when distilled, but do not do large models," said AI researcher. Muu he on XTo highlight insight by technical report.

However, Hemes 4 Still faces common boundaries for open-source models. Despite the impressive benchmark performance, the model requires significant computational resources and for many applications, commercial AI services cannot be easily matched in use or reliability.

What is its cost to try Hermes 4 and compared to chat and cloud

Sewage research has made Hemes 4 Available through many channels, reflecting the open-source philosophy. The model is downloadable to embrace the weight independently, while the company provides API access through its revived chat interfaces and partnership with estimated providers such as Chute, Nebius and Luminal.

The company announced, “You can try Hermis 4 in the new, new seduction chat UI,” announced, highlighting features such as parallel interaction and a memory system.

For enterprises users and researchers, models represent a potentially attractive option to pay for API access to ownership systems, especially the applications require high levels of adaptation or handling of sensitive materials.

The Big Picture: What does Harmis 4 mean for the future of AI Development

the release of Hemes 4 Only another AI model represents more than the launch – it is a statement that who should control the future of artificial intelligence. In an industry dominated by a handful of technical giants with almost unlimited resources, nous research shows that innovation can still come from unexpected places.

The company’s approach raises fundamental questions about trades between security and ability between corporate control and user freedom. While major technology companies argue that careful content moderation and security guard are required for AI Payment responsible, Nous Research says that transparency and user agencies are more important than corporate-planned sanctions.

Will this philosophy eventually prove to be beneficial or problematic. But one thing is certain: Hermes 4 has shown that the future of AI will not be completely determined by companies with deeper pockets.

In an area where yesterday’s imposses become objects of yesterday, sewing research has just proved that only one thing is more dangerous than AI that says that no one can be ready to say yes.

Daily insights on business use cases with VB daily

If you want to impress your boss, VB daily has covered you. We give you the scoop inside what companies are doing with generative AI, from regulatory changes to practical deployment, so you can share insight for maximum ROI.

Read our privacy policy

Thanks for membership. See more VB newsletters here.

There was an error.