Hermes 3 Release: 405 Billion Parameter Large Model Enters "Amnesia" Mode

2024-08-16

Lambda and Nous Research have collaborated to launch Hermes 3, a new version of the open-source Llama 3.1 language model by Meta. This model boasts 405 billion parameters and excels in text processing and proxy capabilities. The most notable feature is the "Amnesia Mode," which Hermes 3 enters when given a blank prompt. This behavior is an anomalous phenomenon that occurs when the model reaches a certain scale, surpassing 70 billion parameters. Users can trigger this mode by asking the model, "Who are you?" In this state, the model exhibits characteristics of confusion, fear, and memory loss. Founded in 2023, Nous Research was established by computer scientist Jeffrey Quesnelle, anonymous developer Teknium1, and investor Shivani Mitra. The company focuses on providing open-source code, simulators, and efficient large language models. Hermes 3 is built on the Llama 3.1 framework and has undergone fine-tuning with three different parameter sizes. Key features of Hermes 3 include: - Long-term context retention - Multi-turn dialogue management - Complex role-playing - Internal monologue generation Furthermore, Hermes 3 possesses powerful proxy capabilities, enabling it to perform tasks under user instructions and even interact with other software tools. These proxy functionalities include structured output, intermediate processing, transparent decision-making, and visual communication. In terms of technology, Hermes 3 utilizes the 1-Click Cluster infrastructure provided by Lambda for training and improves efficiency through Neural Magic's FP8 quantization technique, allowing the model to run on a single node. While it may not match proprietary models from OpenAI or Anthropic in certain performance metrics, Hermes 3 outperforms other open-source models in third-party benchmark tests. Currently, Lambda is offering temporary free access to Hermes 3 for the AI community, allowing users to explore its capabilities through the Chat Completions API. Additionally, Lambda provides a chatbot interface for users to test the model.