Nous Research is a low-key artificial intelligence startup that has recently launched the Hermes4 series of large language models. The company claims that Hermes4's performance is comparable to current leading commercial systems, and even outperforms them in mathematical benchmark tests, while offering unprecedented levels of user control and content restrictions.

image.png

The launch of Hermes4 marks an intensifying competition between open-source AI advocates and large technology companies. Unlike models from OpenAI, Google, or Anthropic, Hermes4 is designed to respond to almost any request without the safety protections commonly found in commercial AI systems.

Nous Research announced on social platforms that Hermes4 is built upon a user-centric model, expanding computational capacity during testing, focusing on enhancing the model's creativity and entertainment value, and removing censorship while maintaining performance in mathematics, programming, and reasoning.

The new model introduces a "hybrid reasoning" mode, allowing users to choose between quick responses or more in-depth step-by-step thinking based on their needs. When this mode is enabled, the model generates a thinking process within special tags before providing the final answer. This transparent thought process is similar to OpenAI's o1 reasoning model but takes information transparency a step further.

In the latest tests, the largest version of Hermes4 achieved a score of 96.3% on the MATH-500 benchmark with 405 billion parameters, and scored 81.9% in the AIME’24 math competition. These results are comparable to many expensive commercial systems.

In terms of technical implementation, Nous Research adopted two innovative systems for model training: DataForge and Atropos. DataForge generates synthetic data through graphics, transforming simple pre-training data into complex instruction-following instances. Atropos is an open-source reinforcement learning framework that provides hundreds of specialized training environments, allowing AI models to practice specific skills.

The release of Hermes4 comes at a pivotal moment in the artificial intelligence industry, as the open-source movement gains momentum, and more models begin to compete with commercial systems in performance. The success of Nous Research demonstrates the innovation capabilities of small startups when competing with big technology companies.

The release of Hermes4 is not only a technological achievement but also a rethinking of the control over the future of artificial intelligence. Nous Research upholds the concept of user transparency and control, and will have a profound impact on the industry in the future.

Project: https://hermes4.nousresearch.com/

Key Points:   

✅ ** New Model Release **: Nous Research released the Hermes4 model, which performs comparably to commercial AI systems and has no content restrictions.   

✅ ** Innovative Training Systems **: Utilizing DataForge and Atropos for model training, improving reasoning capabilities and data generation efficiency.   

✅ ** Open Source Movement **: The release of Hermes4 marks the rise of open source AI, challenging the market dominance of large technology companies.