IT Home reported on November 11 that AI startup Nous Research recently launched an AI chatbot called Nous Chat, which is based on the "Hermes 3-70B" model fine-tuned by meta Llama 3.1, claiming to have made in-depth improvements in "model architecture, data synthesis, fine-tuning and inference", and is committed to "people-oriented".
It is reported that the Nous Chat robot has functions such as long-vein memory, multi-round dialogue, complex role-playing and self-monologue, and at present, Nous Research has also opened the relevant Hermes 3-70B model family in Hugging Face for developers to preview, and the address of the project is as follows (click here to visit).
Nous Research also teased a model called Nous-Hermes-2-Mixtral-8x7B-DPO, an expert hybrid model based on Mixtral 8x7B, which claims to be able to provide a multi-character experience by using the vast amounts of data generated by GPT-4 and other open-source datasets for training.
In addition, the company also announced a Nous-Capybara-7B V1.9 based on the Mistral architecture, which claims to be trained on 20,000 dialogues, which has great advantages in the naturalness and coherence of dialogues.
In terms of Nous Chat, IT Home learned that the overall interface of this chatbot is not much different from ChatGPT and Claude, and it can communicate in English, Chinese, Japanese and other languages, and the knowledge deadline for the corresponding chatbot is April 2023, but the chatbot does not have the programming ability of mainstream AI models, and there is still room for improvement.