Three ways to deploy your own ChatGPT

Original link:


Three Ways to Deploy Your Own ChatGPT – Mr. Fuqiang said: the thinking and precipitation of an architect

Three ways to deploy your own ChatGPT

Wang Fuqiang


What? You still don’t know what ChatGPT is? Well, that’s okay, after all, OpenAI has banned both the mainland and Hong Kong (or not open the service), and now many people can only use it through various tricks. Excusable 🙂

If you really want to experience ChatGPT, but don’t know how to open a channel, create an account or even let it take your money, you can also consider building your own self-hosted chatgpt. Let me introduce three ways to build your own ChatGPT…


LLaMA 1 is a large language model (LLM, Large Language Model) open sourced by Meta company (that is, the original facebook company). The community built a project based on it, called dalai 2 , so that we can easily carry out local It doesn’t matter if you don’t have a high-end GPU, and ordinary notebooks can run:


There is even a project called alpaca-lora 3 , which allows you to run LLaMA on the Raspberry Pi.

But to be honest, after Mr. Fuqiang tried it, he felt that the effect was still somewhat different from the real ChatGPT.

Fortunately, we have a second way…


LLaMA is generally called “Rhana”, although the pronunciation is very similar to a certain monk


Alpaca 4 is “Alpaca”, which is a large language model trained by Stanford University based on the LLaMA model combined with 175 instruction tasks (instruction tasks) for fine-tuning training:


The effect is much better than LLaMA, but, you know, it is an English corpus…

It is easy to deploy by yourself, the open source community is so powerful, someone built alpaca.cpp 5 , running on a single machine, there is no problem at all.

3 ChatGLM

ChatGLM 6 is a Chinese-English bilingual model developed by Tsinghua University. If the previous part is optimized for the English language department, then ChatGLM focuses on the training of Chinese predictions. Run a low-end version on this machine, and the memory only needs 6G.

For the self-deployment large language model of the Chinese scene, ChatGLM is definitely worth a try!

4 words

Of course, in terms of effect, the above three models are actually not as good as ChatGPT. Even Google’s new Bard is said to be unsatisfactory. Therefore, if conditions permit, we still recommend ChatGPT as the first choice. If you are really stuck, then try to deploy one that belongs to Your own big language model is also a good choice.


  1. ↩

  2., this name is a bit sensitive in mainland China… ↩

  3. ↩

  4. ↩

  5. ↩

  6. ↩

Subscribe to "Fu Bao Premium Subscription"

© Fuqiang Wang Personal Copyright, All Rights Reserved.
Copyright © Fuqiang Wang All rights reserved – Since 2004

Everything is homebrewed with pandoc and Markdown , little Scala also included.

This article is reproduced from:
This site is only for collection, and the copyright belongs to the original author.