Exploring OpenChatKit: The Open-Source Alternative to ChatGPT

Introduction

OpenChatKit is an open‑source alternative to ChatGPT released by Together Computer on March 10, 2023. It aims to give developers more control over conversational AI by providing a fine‑tuned large language model, modular components, and a permissive Apache-2.0 license.

What Is OpenChatKit?

Developed by former OpenAI researchers, OpenChatKit bundles a suite of tools for building general‑purpose or specialized chatbots. The project is hosted by Together, a startup focused on decentralized AI cloud services, and includes source code, model weights, and training data.

Key Components

Instruction‑tuned LLM based on EleutherAI’s GPT‑NeoX‑20B, fine‑tuned on 43 million high‑quality dialogue instructions.

Documentation describing how the fine‑tuned model achieves high accuracy on specific tasks.

A scalable retrieval system that can augment responses with up‑to‑date knowledge from Wikipedia, news feeds, or sports scores.

A moderation layer fine‑tuned from GPT‑JT‑6B to filter inappropriate outputs.

Model Architecture

The core model, called GPT‑NeoXT‑Chat‑Base‑20B , builds on the GPT‑NeoX architecture and has been evaluated on the HELM benchmark, where it outperforms its base model.

Limitations

Knowledge: The chatbot may produce factually incorrect answers, similar to ChatGPT.

Code: It has not been trained on a large code corpus, limiting its ability to generate accurate code.

Context switching: It can fail to maintain topic continuity when the conversation shifts.

Repetition: Responses may repeat or stall, requiring a page refresh.

Creativity: It tends to give short answers and does not generate long‑form creative text.

Conclusion

OpenChatKit provides a solid starting point for developers who want an open‑source conversational model. While it is still early‑stage and less capable than proprietary systems like ChatGPT, community contributions are rapidly improving its performance.