Hermes 3: The First Fine-Tuned Llama 3.1 405B Model

Nous Research has launched Hermes 3, the first full-parameter fine-tune of Meta's Llama 3.1 405B model, trained on Lambda's 1-Click Cluster. Hermes 3 is a neutrally-aligned, generalist model with strong reasoning capabilities, designed for the open-source community and available for free via Lambda's Chat Completions API. The model excels in creative tasks like role-playing and fiction, as well as in professional applications requiring advanced reasoning and decision-making. Hermes 3 was trained using synthesized data, supervised fine-tuning, and reinforcement learning from human feedback, followed by Neural Magic’s FP8 quantization, reducing its VRAM and disk requirements by 50%. It can run efficiently on a single node or scale to a multi-node cluster for further fine-tuning. Hermes 3 is unlocked, uncensored, and steerable, providing flexibility and alignment with user needs. The model outperforms Llama 3.1 Instruct on benchmarks and is available for free through Lambda’s new Chat Completions API, which is compatible with the OpenAI API. The API offers easy access with no complex setup, allowing users to generate completions and chat completions effortlessly.