By Wang Zhiyuan | ID:Z201440 By Wang Zhiyuan | ID:Z201440

The Qwen3 series models from Alibaba Cloud have been the talk of the town for a while.

Yesterday evening, several friends around me were waiting for the release, and I was watching a show while waiting. I was so tired that I fell asleep at 1 a.m.

This morning, I woke up early and saw the official announcement. My colleagues inside said the model went live at 4 a.m. yesterday, and the official announcement came out around 6 a.m.

These people are really dedicated; they worked through the night to get it done before May Day. They didn't sleep.

Officially, this Qwen3 has made significant improvements in natural language processing and multi-modal capabilities, marking a new upgrade in Alibaba Cloud's artificial intelligence domain.

Its popularity doesn't need much explanation; within two hours of release, the number of stars on GitHub exceeded 16.9k. The entire series includes 2 MoE models and 6 dense models, making it quite impressive.

What are MoE models? And what are dense models?

You can think of MoE models as a council of "experts," each responsible for their own area. When you have a complex problem, it will determine which expert is best suited to handle it and then call upon that expert to help solve your issue.

Dense models are straightforward; they are all-rounders capable of handling any task independently. Their advantage lies in not needing to consider division of labor; they just dive in. For less complex problems, they perform well.

However, if the problem is too complex or specific, it may not be as professional, after all, they are not designed for specialized fields.

The models released this time can basically be experienced in the model market, Alibaba Cloud's official website, huggingface, or the ModelScope community. Detailed parameters are not listed here, but they can easily be found with an online search.

For ordinary users who want to try it themselves, the Tongyi app on mobile devices is sufficient.

I found when I opened the app this morning that three models had already been launched: one called A22B, another named 32B, and another high-performance deep-thinking version of 32B.

The "B" here stands for "Billion," indicating the number of model parameters.

For example:

32B means 32 billion parameters, and A22B can be understood as having a parameter scale of around 22 billion. So, when a model is followed by how many Bs, it’s essentially telling us how "big" it is. More parameters generally mean stronger model capabilities.

As I wrote this, a question arose: Which one should I choose for daily use?

Qwen3-235B-A22B is too large for most everyday uses; it would be suitable for developers and enterprises. I feel it would be a waste to run it on a mobile phone.

Qwen3-32B is more suitable for daily use; it can act as an assistant for handling documents, chatting, writing simple reports, etc., and falls into the category of medium-sized models.

QwQ-32B, on the other hand, is a "thinking" model, ideal for situations where you're overwhelmed and don’t want to think too hard. For instance, if you want to make a PowerPoint presentation, you could ask it to first outline the framework, then refine it gradually.

So, I think Qwen3-32B and QwQ-32B are enough.

I tried these two models on my iPhone 14 Pro Max. The running speed was very fast, with no lagging, and the phone did not overheat. The text generation speed was clearly faster than my speech.

To be honest, I rarely use AI models on my phone because I mostly work on PC.

Speaking of PC, the web address is: https://chat.qwen.ai

I counted, and there are more than 10 models here.

These models are some of the commonly used and larger models under Alibaba Cloud. Many small models derived from each of these models are widely used by developers on GitHub.

Therefore, isn't it a great experience to directly converse with these models on a PC?

Previously, I often used the Qwen 2.5 Max model, which was particularly sensitive in content processing and could accurately understand my intentions. After this update, I first tried its A22B version. It felt uninteresting to let it process text, so I asked it to generate an image instead.

I had Kimi help me write an evocative prompt, then copied the prompt to A22B. It reacted quickly, generating an image in about three seconds, sized at 1.5 MB. What do you think?

This model supports continuous dialogue, but there seems to be a problem with its continuous dialogue capability: it can only generate new images and cannot optimize existing ones.

This might be because the model itself hasn't been optimized for continuous reasoning based on context in chat boxes. However, it's not a big deal.

I uploaded a video I took the day before, about the Quark AI Super Search Box, which was approximately 210MB and lasted 2 minutes, and then processed it using the 30B-A3B model.

I told it: "Help me extract the text from the video, optimize it, and I'll post it on Xiaohongshu (Little Red Book)."

In fact, this is somewhat of a complex or sequential task that requires multiple steps. But surprisingly, it extracted the text for me and even came up with a title that looked perfect for posting on Xiaohongshu.

The whole process went smoothly, and the efficiency was high.

To be honest, I can't describe its speed and efficiency in words or examples anymore. It feels like its processing ability far exceeds the normal functioning of my brain.

My limitation lies in not knowing how to use it effectively. Or rather, facing such a powerful tool, I wonder:

Where is my upper limit of thinking? What other complex tasks can it help me solve? In daily life, what else can I entrust to it?

I started to worry that I might gradually become dependent on it or even be led by it.

Aristotle once said: "Human nature lies in seeking knowledge." Thinking used to be important, but now it seems less "necessary." If so, what is truly important?

Perhaps, we need to reconsider the topic of "how to coexist with AI."

Original article: https://www.toutiao.com/article/7498544115865354791/

Disclaimer: This article solely represents the author's personal views. Please express your opinions in the buttons below to "like" or "dislike" this article.