A Review of Major AI Models in China and Abroad

As a heavy user of productivity tools, I consider myself to be on the cutting edge of this wave of AI. I have basically used various AI models and tools, and here I will briefly share my views on them. The following evaluations are very subjective, and I don’t care about any benchmark rankings. I will discuss them based on actual usage experience.

🏠 Let’s Start with Domestic Models:

The most popular domestic models now are Kimi, Zhipu Qingyan, Tongyi Qianwen (Alibaba), Doubao (ByteDance), and Lingyi Wanyi, etc. These few are worth mentioning; other domestic models generally lack significant highlights, including Baidu’s Wenxin Yiyan.

I mainly use Kimi and Zhipu Qingyan among domestic models; I have little interest in the others. From what I understand, domestic models are mostly similar, based on Meta’s (Facebook’s) open-source LLaMA models or modified GPT-2, with their own Chinese training data. However, even in Chinese capabilities, they still cannot match GPT-4.

Kimi: This model mainly supports around 2 million context tokens in Chinese. GPT-4 only supports 32K context tokens, so you can directly throw in an eBook for analysis, which is quite enjoyable.

Zhipu Qingyan: This model mainly supports customizable bots, eliminating the hassle of rewriting prompts each time. This feature is only available in the paid version of GPT-4 Plus, and currently, it’s the only domestic model that supports this. The “English Polishing Translation Master” bot I created is also on this platform. Additionally, their latest GLM-4 model has indeed been tested to be slightly stronger than GPT-3.5.

Tongyi Qianwen: Alibaba’s model ranks quite high on various lists. I have used it a few times but didn’t feel anything particularly special. The most I used from them is the “Tongyi Listening and Understanding,” which is an enhanced speech-to-text function. During phone meetings with foreigners, I can see the text content in real-time, and the response is very quick.

Doubao: Produced by ByteDance, their AI drawing function is currently the strongest domestically, capable of generating realistic-style images:

Additionally, it can have real-time voice conversations. Currently, it doesn’t support much else, but if you set the conversation language to English, it can act as an AI foreign teacher, with about 70% proficiency.

Lingyi Wanyi: Founded by Kai-Fu Lee, it was recently revealed to actually be a shell of Meta’s open-source LLaMA model. He openly admitted that it borrowed the basic architecture of LLaMA, and other so-called self-developed domestic models are likely similar.

The highlight of their model is that it can generate PPTs automatically based on input prompts, complete with detailed content and images. It’s decent, though the template styles are not very aesthetically pleasing; they do match the aesthetic of older domestic leaders, and I occasionally use it.

Finally, I want to mention a domestic AI tool that is not an AI model but is quite useful:

Mitai Search: This tool is comparable to Perplexity AI abroad. When you input a question, it automatically searches the web and summarizes the results into an answer. It can be considered a smart version of Google, very similar to New Bing but with much stronger search capabilities.

Dreamina: An AI video generation tool from ByteDance, currently the strongest domestically. You can generate a few seconds of video just by uploading a photo.

🌍 Now, International Models:

ChatGPT: Undoubtedly the leader among AI models, GPT-3.5 has been surpassed by many models, but GPT-4 remains unmatched to this day. GPT-3.5 is already open for use without registration, while GPT-4 still costs $19.9 per month. The latest generation GPT-4o is slightly stronger than GPT-4 and faster. Free users can only try it a few times before needing to recharge.

Claude: Created by former OpenAI employees Dario Amodei and Daniela Amodei, who themselves were former Baidu employees, Claude is the second-best model after ChatGPT. Its capabilities in literature and language even surpass ChatGPT. I often use Claude for polishing my articles, and their free version Claude-3-Sonnet also exceeds GPT-3.5.

LLaMA: Meta’s open-source model, the latest LLaMA3 has surpassed GPT-3.5 in capability. Many domestic models have traces of LLaMA in them since they are all based on its open-source version. Local deployment is also easy; just download Ollama or use meta.ai online. It answers English content much faster and better than GPT-3.5, but its support for Chinese content is lacking, and it often switches back to English mid-response.

Gemini: Google’s model, in actual use, still feels like it summarizes content from the web. If the information is not online, it cannot answer. Currently, Google’s model is lagging behind in the first tier abroad.

Aside from major models, let me mention a few popular international AI tools:

Midjourney: The strongest AI drawing tool globally, without a doubt. Any other player in this field can only see its tail lights. Their costs are also high; $10 only allows for 200 images.

Perplexity AI: AI search, more powerful than New Bing, automatically searches the web for answers and summarizes the results.