Lambda, an AI infrastructure company, and Nous Research, a startup focused on personalized AI, have collaborated to release Hermes 3, a new version of Meta’s open-source large language model (LLM), Llama 3.1. The most intriguing feature of Hermes 3 is its ability to experience an existential crisis when given a blank prompt. This discovery was unexpected and points to anomalous behavior that emerges when scaling AI models beyond certain thresholds. Users are encouraged to interact with Hermes 3 on Discord to uncover more about its capabilities.
Nous Research, co-founded by Jeffrey Quesnelle, Tanishq Abraham, and Shivani Mitra, raised $5.2 million in seed funding in January 2024. Their mission is to provide potent open-source code and efficient large language models. Hermes 3 follows in the footsteps of its predecessors, Hermes, Hermes 2, and Open Hermes 2.5, which have collectively been downloaded 33 million times. Unlike other models, Hermes 3 offers unlocked and uncensored open weights, allowing users to tailor its responses to their individual needs.
Hermes 3, built on the Llama 3.1 framework, has been fine-tuned across three different parameter sizes: 8B, 70B, and the largest, 405B. It was trained on a diverse dataset to enhance its reasoning, creativity, and adherence to user instructions. Some of its capabilities include long-term context retention, multi-turn conversation management, complex role-playing, and internal monologue generation. Later this year, Nous plans to release an open-source AI orchestration platform called “Nous Forge.”
Hermes 3 stands out for its agentic capabilities, which refer to AI models performing actions on behalf of users. It can use XML tags for structured output, generate internal monologues for transparent decision-making, create visual communication using Mermaid diagrams, and employ step-labeled reasoning and planning. The model showcases proficiency in generating complex, functional code snippets and providing detailed code explanations and documentation, making it valuable for software development and code analysis.
Hermes 3 was trained using Lambda’s 1-Click Cluster infrastructure, which provided remarkable results within a few weeks. The model is optimized for efficiency, with techniques like Neural Magic’s FP8 quantization reducing VRAM and disk requirements by approximately 50%. While not as performant as closed-source models from companies like OpenAI or Anthropic, Hermes 3 outperforms other open-source models in benchmark tests.
Hermes 3 is a versatile tool suitable for various applications. It excels in scenarios that require advanced reasoning, strategic planning, decision-making, and creative storytelling. Lambda offers temporary free access to Hermes 3 through its Chat Completions API, fully compatible with the OpenAI API. Users can generate a Cloud API key via Lambda’s dashboard and test the model’s capabilities without complex setup. Dedicated access to Hermes 3 can be deployed on a single Lambda node or scaled to a multi-node configuration for further fine-tuning.
Lambda and Nous Research are eager for users to engage with Hermes 3 and share their findings. As AI continues to evolve, Hermes 3 represents a user-centric and adaptable model that offers a glimpse into the future of AI.