Original video in Chinese.
Key Takeaway
- HuggingChat is an app that lets you experience the most advanced open-source LLMs for free, with web, iOS, and macOS versions, and a clean design.
- HuggingChat’s core proposition is to let everyone use the best models from the Hugging Face community, with its model list updated from time to time.
- I use the Q&A engine Perplexity and the chatbot Claude very frequently every day, but HuggingChat has become my go-to tool for lightweight, fragmented needs because of its lightness and convenience.
- HuggingChat brings up the chat window through a shortcut, and prioritizes faster replies; the Web Search feature has to be turned on manually.
- HuggingChat also offers a Tools feature, and its Flux image generation tool can meet lightweight image generation needs.
- The article argues that open-source models have now matched closed-source models in performance, and the open-source community has an advantage in building applications.
If you want to experience the most advanced open-source LLMs for free, I recommend giving HuggingChat a try.
This app used to have a web version and an iOS version. A few days ago, the macOS version launched, and there was quite a bit of thought put into the product design. I especially agree with this kind of clean design approach.
The macOS version of HuggingChat is not like other apps, with a heavy frontend. Only after pressing the default shortcut of Command, Shift, and Enter does a minimalist chat window appear, very much like macOS Spotlight Search. At that point, we can talk to a large language model running in the cloud.
If you want to switch models, click the plus sign on the left to go into settings, and change the model from the default Llama 3.1 70B to the domestic Qwen 2.5 72B. These models are not fixed; they are updated from time to time. Because HuggingChat’s proposition is:
Let everyone use the best models from the Hugging Face community.
This also shows that Qwen 2.5, like Llama 3.1 and Command R+, has become a recognized, currently best open-source LLM. Qwen really is a source of national pride!
Hello everyone, welcome to my channel. I’m one of the few creators in China who can explain both the why and the how of AI clearly. Remember to hit follow — you definitely won’t regret it. If you want to connect with me, come join the newtype community; more than 500 friends have already paid to join.
Back to today’s topic: HuggingChat. The AI tools I use frequently every day fall into two categories:
The first is Q&A engines. Right now, the best Q&A engine in the world is Perplexity, no question, no competition. But that alone is not enough, because there are many things you can’t search for. And I often need AI to provide more angles or help me refine my thinking.
That leads to the second category of tools: chatbots. The one I’m most satisfied with right now is Claude. Not only is its model capability stronger than GPT-4’s, but the product experience is also much better than ChatGPT’s. Artifacts are really great — absolutely worth the money. I’ve recommended it to a lot of people, and everyone who’s used it says it’s good.
However, for us users in China, the annoying thing about Perplexity and Claude is that every so often you have to refresh the page and click verification. Don’t underestimate even this tiny bit of trouble — when you want to use it but get interrupted, it really affects the experience.
So after dealing with this inconvenience for a while, I only use those two when I have somewhat more serious tasks. For the everyday, highly fragmented needs, I need a lightweight AI tool that I can pick up and use right away — that’s why I took a liking to the macOS version of HuggingChat.
Normally hidden in the background, and brought up with a shortcut when needed, this seemingly non-confrontational approach is actually trying to seize the first entry point to the AI terminal. To achieve this ambition, HuggingChat has made a lot of cuts, and even the web search feature has to be turned on manually.
There’s a Web Search setting. Once you check it, the model will search the web. The tradeoff is that replies become a bit slower, because there’s an extra search and RAG process. I guess that’s why web search isn’t turned on by default.
Doing everything possible to speed up replies is definitely higher priority than any other feature.
If users have heavier needs, no problem — use the web version on desktop and the iOS version on mobile. Once you open it, you’ll see that it has something like ChatGPT’s GPTs too, called Assistants. But most of them are not very useful, just like GPTs.
What’s truly productive is Tools. The one I use most is Flux image generation.
I introduced the Flux model in the previous two videos. It was created by the SD team, and it is currently the most advanced image generation model in the world. First, the images Flux generates surpass other models both in realism and aesthetic quality. Second, Flux can also achieve precise control, such as accurately generating text within images.
Flux has three versions, two of which are open source. The Flux dev used by this tool is the most capable one among the open-source versions. I use it when I have some lightweight image generation needs. For example, generating an illustration for an article. Since Flux itself is so capable, this kind of task is easy for it. If it doesn’t work the first time, just sample a few more times and it can be done.
After the macOS version came out, combined with the iOS version and web version I was already using, I suddenly realized that HuggingChat had quietly become the AI tool I use most frequently. Open-source models have already caught up with closed-source ones in performance. As for app development, everyone is at about the same level. I even feel that the open-source community has an advantage, because they don’t have to think about ecosystems or moats or things like that. They have fewer burdens and can just let loose and do the work.
OK, that’s it for this episode. If you want to discuss and learn AI, come join the newtype community. See you next time!