8 Open-Source ChatGPT Alternatives You Can Deploy Today

This article introduces eight widely used open-source frameworks for building ChatGPT-like chatbots, integrating LLMs, or launching new AI software products.

LLaMA

The LLaMA project provides a suite of foundational language models ranging from 7B to 65B parameters, trained on millions of tokens from publicly available datasets. LLaMA‑13B outperforms GPT‑3 (175B), and LLaMA‑65B matches the performance of models such as Chinchilla‑70B and PaLM‑540B.

Resources:

GitHub: https://github.com/facebookresearch/llama

Demo: https://huggingface.co/spaces/project-baize/baize-lora-7B

Alpaca

Stanford Alpaca claims to compete with ChatGPT and can be reproduced for under $600. Alpaca‑7B is fine‑tuned from LLaMA‑7B on 52K instruction data.

Resources:

Blog: https://crfm.stanford.edu/2023/03/13/alpaca.html

GitHub: https://github.com/tatsu-lab/stanford_alpaca

Demo: https://huggingface.co/spaces/tloen/alpaca-lora

Vicuna

Vicuna is fine‑tuned from LLaMA using user‑shared dialogue data collected from ShareGPT. Vicuna‑13B reaches over 90% of the quality of OpenAI ChatGPT and Google Bard, often surpassing LLaMA and Stanford Alpaca. Training costs about $300.

Resources:

Project page: https://vicuna.lmsys.org/

GitHub: https://github.com/lm-sys/FastChat

Demo: https://chat.lmsys.org/

OpenChatKit

OpenChatKit is a complete toolkit for building chatbots. It provides instruction‑tuned large language models, fine‑tuning pipelines, a scalable retrieval system for updating responses, and moderation instructions for filtering queries.

Resources:

GitHub: https://github.com/togethercomputer/OpenChatKit

Article: OpenChatKit: Open‑source ChatGPT alternative

GPT4All

GPT4All is a community‑driven open‑source project trained on a curated corpus of code, stories, and descriptions. The team provides datasets, model weights, data pipelines, and training code, including quantized models that run on notebooks. A Python client enables inference.

Resources:

Technical report: https://s3.amazonaws.com/static.nomic.ai/gpt4all/2023_GPT4All_Technical_Report.pdf

GitHub: https://github.com/nomic-ai/gpt4all

Demo (unofficial): https://huggingface.co/spaces/rishiraj/GPT4All

Model: https://huggingface.co/nomic-ai/gpt4all-lora

Raven RWKV

Raven RWKV‑7B is an open‑source chatbot powered by the RWKV language model, delivering ChatGPT‑like answers. Its RNN architecture matches ChatGPT in quality and scaling while being faster and using less VRAM. It has been fine‑tuned on Stanford Alpaca, code‑alpaca, and other datasets.

Resources:

GitHub: https://github.com/BlinkDL/ChatRWKV

Demo: https://huggingface.co/spaces/BlinkDL/Raven-RWKV-7B

OPT

OPT (Open Pre‑trained Transformer) lags behind ChatGPT in overall performance but excels at zero‑shot, few‑shot learning and bias analysis. It can be integrated with Alpa, Colossal‑AI, CTranslate2, and FasterTransformer for improved results. It is popular, with over 624,000 monthly downloads.

Resources:

Paper: https://arxiv.org/abs/2205.01068

GitHub: https://github.com/facebookresearch/metaseq

Demo: https://huggingface.co/spaces/tomg-group-umd/lm-watermarking

Model hub: https://huggingface.co/facebook/opt-1.3b

Flan‑T5‑XXL

Flan‑T5‑XXL is an instruction‑tuned T5 model fine‑tuned on over 1,000 tasks, improving performance of models such as PaLM, T5, and U‑PaLM across many languages.

Resources:

GitHub: https://github.com/google-research/t5x

Demo: https://huggingface.co/spaces/olivierdehaene/chat-llm-streaming

In summary, these open‑source chatbots and models are rapidly improving, and a future model may surpass ChatGPT in performance.