Nous Research Unveils DeepHermes: New Open AI Models with Toggleable Reasoning Capabilities

Nous Research has announced the release of its latest AI models, DeepHermes 24B and 3B Preview, introducing what they call "Hybrid Reasoners" to the open-source AI ecosystem. The models were unveiled on March 13 via the company's social media channels and are now available for download on Hugging Face and through the Nous API.

"Our best hybrid reasoner is now available! DeepHermes 24B is built on @MistralAI's Open 24B Mistral-Small model and is a real beast. We also released a new, smaller 3B DeepHermes for low resource edge reasoning!" said Teknium, co-founder of Nous Research. "I am incredibly proud of how good DeepHermes 24B is at both objective tasks and less verifiable tasks!"

What sets these models apart is their ability to toggle between reasoning modes. Users can switch between getting quick, intuitive answers or more detailed, well-reasoned responses with higher accuracy. This flexibility makes the models adaptable to different use cases, from casual queries to complex problem-solving.

According to Nous Research, the DeepHermes 24B Preview shows remarkable performance on reasoning tasks when in reasoning mode, with over four times higher accuracy on difficult math problems and a 43% improvement on GPQA, a benchmark for STEM-based question answering. Built on Mistral AI's Mistral-Small-24B open model, it's designed to be efficiently quantized for consumer GPUs, making it accessible for local deployment.

The company notes that the models scale effectively with size, with performance progressively improving from the 3B to the 24B version. Beyond objective tasks, Nous Research claims the models excel at questions requiring deep thought while maintaining transparency in their reasoning process.

DeepHermes was developed by a team of AI researchers including Teknium, Roger Shi, JSupa15, nullvaluetensor, and theemozilla, along with contributions from the open-source community worldwide.

For those interested in trying the models, Nous Research has created a free interactive DeepHermes 24B chatbot in their Discord server at https://discord.gg/NousResearch.

For developers and researchers looking to use or deploy the models:

• The complete DeepHermes collection is available at: https://huggingface.co/collections/NousResearch/deephermes-67d2ff8c9246cc09a7bd8add
• API access (waitlist): https://portal.nousresearch.com
• Download the 24B model: https://huggingface.co/NousResearch/DeepHermes-3-Mistral-24B-Preview
• Download the 3B model: https://huggingface.co/NousResearch/DeepHermes-3-Llama-3-3B-Preview
• GGUF quantized version (24B): https://huggingface.co/NousResearch/DeepHermes-3-Mistral-24B-Preview-GGUF

This release continues Nous Research's commitment to developing and sharing open-source AI models and follows their previous work on the Hermes series of language models.