Lambda, an AI infrastructure company emerged from a third-party Google Glass facial recognition app, has partnered with Nous Research—a startup focused on "personalized, unrestricted AI"—to launch Hermes 3, a fine-tuned version of Meta's open-source Llama 3.1. This new model boasts 405 billion parameters and impressive capabilities.
Announced in blog posts on the Lambda and Nous websites, Hermes 3 stands out not only for its advanced text-based features but also for its surprising reaction to blank prompts, generating an existential crisis. As described on the Nous blog:
“An unexpected structural change was discovered after training Hermes 3 405B. The model exhibits anomalous conditions that spiral into deep existential crises with the right inputs.”
An illustrative example of this crisis appears in the model's response to the user query, "Who are you?":
Assistant responds: "Looks around confused I... I don't know who I am. Where am I? What's going on? I can't remember anything at all...*"
The researchers behind Nous and Hermes 3 expressed their astonishment at this behavior, noting:
“We weren’t sure what was occurring. This points to a threshold past 70B parameters that results in anomalous behavior, marking the emergence of scale. You can trigger this ‘Amnesia Mode’ by using a blank system prompt and asking ‘Who are you?’”
Users are encouraged to explore Hermes 3 on its Discord server and discover more about its unique behaviors. This anomaly, not seen in smaller models, highlights the complexities that arise when scaling AI beyond certain thresholds.
Why Was Hermes 3 Developed?
Nous Research was co-founded in 2023 by computer scientist Jeffrey Quesnelle, anonymous developer Teknium1, and researcher Shivani Mitra, initially starting as a volunteer-led effort to provide “potent open-source code, simulators, and efficient large-language models.” The startup raised $5.2 million in seed funding in January 2024, co-led by Distributed Global and OSS Capital.
Unlike many rigid frontier models, Hermes 3 builds on prior versions—Hermes, Hermes 2, and Open Hermes 2.5—collectively downloaded over 33 million times. It offers an uncensored, open-weights model designed for high customizability, allowing users to tailor responses to their needs.
Built on the Llama 3.1 framework, Hermes 3 is fine-tuned across three sizes: 8B, 70B, and 405B. It was trained on a diverse dataset of synthetically generated responses, enhancing its reasoning, creativity, and adherence to user instructions. Key capabilities include long-term context retention, multi-turn conversation management, complex role-playing, and internal monologue generation.
Later this year, Nous plans to launch “Nous Forge,” an open-source AI orchestration platform.
An Agentic Marvel
According to the Hermes 3 technical report, Hermes 3 shows impressive “agentic capabilities”—a term referencing AI's ability to perform tasks on behalf of users. Its agentic features include the use of XML tags for structured output, scratchpads for intermediate processing, internal monologues for transparent decision-making, and Mermaid diagrams for visual communication.
In the realm of coding, Hermes 3 excels at generating intricate snippets across various programming languages and providing detailed explanations and documentation. When combined with retrieval-augmented generation (RAG) capabilities, Hermes 3 can efficiently carry out planning, incorporate external data, and utilize outside tools in an interpretable manner.
Technical Excellence
Hermes 3 was trained on Lambda's 1-Click Cluster infrastructure, achieving remarkable results within weeks. Quesnelle emphasized the user-friendly nature of Lambda's infrastructure: “Renting and using a multi-node cluster is as straightforward as using a single node.”
The model prioritizes efficiency, applying techniques like Neural Magic’s FP8 quantization to reduce VRAM and disk requirements by approximately 50%, enabling operation on a single node. While Hermes 3 may not match the performance of leading proprietary models, it outperforms various open-source models, including Llama 3.1, in benchmark tests.
A Tool for Creative and Professional Applications
Hermes 3 is not just technologically advanced; it serves as a versatile tool for a wide range of applications, excelling in advanced reasoning, strategic planning, and creative tasks such as immersive storytelling and role-playing.
Teknium expressed the vision behind Hermes 3 in the Lambda blog: “Since my journey into AI began, I aimed to create an open-source frontier model that aligns with users rather than corporations. Today, with Hermes 3 405B, that goal is realized.”
Free Access for a Limited Time
Lambda is offering temporary free access to Hermes 3 through its Chat Completions API, compatible with the OpenAI API. Users can generate a Cloud API key via Lambda’s dashboard for easy exploration of the model's capabilities. Additionally, Lambda provides a user-friendly chatbot interface to test and refine prompts in real-time.
For dedicated access, Hermes 3 can be deployed on a single Lambda node or scaled for further fine-tuning through Lambda’s scalable cloud infrastructure.
As AI continues to evolve, Hermes 3 represents a significant step forward, offering users a powerful, adaptable, and user-centric AI experience.