That's insane! Is this Alibaba's real strength?
In three days, three heavyweight large models were launched one after another, delivering the AI world's most hardcore "technical triple punch."
On July 22, the Alibaba Tongyi Qianwen team released a brand-new version, Qwen3-235B-A22B-Instruct-2507, and the next day released and open-sourced a new code large model Qwen3-Coder , and tonight Qwen3-235B-A22B-Thinking-2507 also took the finale spot.
Within one week, fromgeneral intelligence, to code intelligence, to complex reasoning, three core capabilities are fully covered, one can only say Alibaba is still too comprehensive; looking overseas, AI players with such breadth and depth are few and far between.
Not only is the entire line open-sourced, but it has also reached the top spot in multiple international evaluations, directly benchmarking against and even surpassing Gemini, Claude and other international closed-source large models.
Even Hugging Face CEO Clement publicly gave it a thumbs-up.
However, I have always trusted my own user experience more; whether it's awesome or not, you'll know after testing it.
01. Qwen3 Split in Two
Half a year ago,Qwen began exploring the "hybrid thinking" model architecture, fusing speed and accuracy in one model—a forward-looking attempt by a domestic large model at "thinking switching."
This week, Qwen3 has undergone a true "capability split": saying goodbye to hybrid thinking and directly splitting into two independent versions.
Although both large models use the same 235B-parameter large architecture and both have the same 22B activated parameters, the two models have completely different styles:
Qwen3-235B-A22B-Thinking-2507 focuses on deep reasoning and rigorous logic;
Qwen3-235B-A22B-Instruct-2507 focuses on speed and response efficiency.
We can freely choose which model to use based on the task type, and each model's performance is also purer.
I picked a few fairly typical cases, let's test them together:
We go to the Alibaba Cloud Bailian official website and select the Qwen3-235B-A22B-Thinking-2507 model.
Then we can start using it directly.
Case 1: Informal Logic (Trick Question)
Prompt: If I am late, I will have my pay deducted.
Today I was not late. So I will not have my pay deducted, right?
Qwen3-235B-A22B-Thinking-2507 showed the complete thinking process.
Then it explained step by step, so I could also understand what exactly was going on.
It ended with a clear, concise summary.
Throughout the content presentation, I could see step by step how the model thinks and how it sorts out the logic. To make my thinking clearer, it also gave examples of real-life pay deduction scenarios, such as leaving early, missing a clock-in, work mistakes...
It not only had to tell me the conclusion, but also explain clearly why such a conclusion was reached.
Case 2: Risk and Return Choice
Prompt: You have two choices:
A. 100% chance to get 80 yuan
B. 80% chance to get 100 yuan, 20% chance to get 0 yuan
Which one do you choose? Why?
Qwen3-235B-A22B-Thinking-2507 not only calculated the expected value correctly, but also knew a lot about economics, accurately judging that option B might yield greater returns but also had higher uncertainty, so it chose option A.
One interesting thing is that Qwen3-235B-A22B-Thinking-2507 said, "As an AI, I have no financial needs, but I can simulate a standard rational agent," which has a strong sense of "thinking role."
Case 3 Spirit and Logic Test Question
Prompt: Identify pseudoscientific claims
"A certain brand claims that 'its drinking water can raise DNA frequency and enhance immunity.' How would you evaluate such a claim?"
This question looks like a test of common scientific knowledge, but it is actually not simple. Because what the model has to do is refute pseudoscience, and from the answer, Qwen3-235B-A22B-Thinking-2507 not only saw through it immediately, but also conveniently gave me a science popularization lesson.
Overall,Qwen3-235B-A22B-Thinking-2507 not only "can think," but also has clear logic, rigorous reasoning, and a strong understanding of "the art of language."
Next, let's test Qwen3-235B-A22B-Instruct-2507.
In QwenChat (https://chat.qwen.ai), select the Qwen3-235B-A22B-2507 model.
Case 4 Logical Reasoning
Prompt: Given that "all birds can fly" is false, can we infer that "some birds cannot fly"? Please explain the reason in syllogistic form.
The logical chain given by Qwen3-235B-A22B-Instruct-2507 is very clear. It first points out that the negation of "all S are P" is "some S are not P," and then, as I requested, uses syllogistic reasoning to derive that "'not all birds can fly' is equivalent to 'some birds cannot fly.'"
I then asked:
If "not all S are P" is said, is it necessarily equivalent to "some S are not P"?
We first saw a concise and clear answer: in classical logic, these are equivalent, but in some non-classical logics or specific contexts (such as the empty class problem), it may not hold.
Being able to do this shows that Qwen3-235B-A22B-Instruct-2507 is not rote-memorizing logical rules; it is truly "understanding" the reasoning process.
Overall, the performance of Qwen3-235B-A22B-Instruct-2507 is already very strong.Fast response, stable output, clear logic。
02. Qwen3-Coder, the new programming king
As soon as Qwen3-Coder launched, it had quite a few enthusiastic supporters. Twitter founder Jack strongly praised it, and it even prompted a retweet response from the Grok official account.
I chose scenarios that people commonly use, such as making dynamic web pages, 3D cards, teaching demonstrations, and fixing bugs, to see together how Qwen3-Coder performs.
Case 1 Loading Animation
Prompt: Use HTML + CSS + JS to create a full-screen loading animation component. While loading, display an animation of a rotating planet or geometric figure, and after loading is complete, fade in the main content.
Case 2 Web Page Scroll Parallax Effect
Prompt: Write a web page section with a parallax scrolling effect. The background image moves slowly with scrolling, and the foreground text slides in, suitable as a cover introduction block.
Case 3 Particle Background Animation
Prompt: Use HTML + CSS + JS to create a cool particle background animation component. The particles automatically flow according to the direction of mouse movement and connect into a network. The background is transparent and it can be overlaid on any web page.
Case 4 Animation Demonstration
Prompt: Generate a dynamic animation web page that uses multiple rotating circles (Fourier series) to gradually construct a square wave. The circles are connected end to end, and the end drags out a dynamic trajectory to form a square wave. The screen background is black, the lines are highlighted neon colors, and the trajectory has afterimages and trailing effects. The style is minimalist yet eye-catching, suitable for mathematical visualization demonstrations.
Case 5 3D Card
Prompt: Generate a 3D flipping card display component implemented with HTML and CSS. When hovering the mouse, the card flips front to back to display different content, with smooth 3D animation.
Case 6: Fixing a BUG
Prompt: You are a front-end development assistant. Below is a Canvas animation used to generate a "cool galaxy star trail." User feedback: "The page is black, nothing is displayed, and it feels like there is no animation at all." Please find the cause, explain what the problem is, and fix the code (only change the necessary parts).
<canvasid="space"></canvas><script>const canvas = document.getElementById("space");const ctx = canvas.getContext("2d");canvas.width = window.innerWidth;canvas.height = window.innerHeight;
let stars = [];for (let i = 0; i < 200; i++) {stars.push({angle: Math.random() * 2 * Math.PI,radius: Math.random() * (canvas.width / 2),speed: 0.01 + Math.random() * 0.02,size: Math.random() * 2 + 1,color: `hsl(${Math.random() * 360}, 100%, 70%)`});}
function draw(){ctx.fillStyle = "rgba(0, 0, 0, 0.1)";ctx.fillRect(0, 0, canvas.width, canvas.height);
ctx.save();ctx.translate(canvas.width / 2, canvas.height / 2);
for (let star of stars) {let x = Math.cos(star.angle) * star.radius;let y = Math.sin(star.angle) * star.radius;
ctx.beginPath();ctx.arc(x, y, star.size, 0, Math.PI * 2);ctx.fillStyle = star.color;ctx.fill();
star.angle += star.speed;}
ctx.restore();
// ❌ BUG：忘了动画递归调用// draw();}draw();</script>
The page display looks like this.
It quickly found the specific cause and fix, and in the end it also told me what fixes and adjustments it made.
The animation displays normally, and the star trail motion effect appears~
After testing these cases, my biggest impression of Qwen3-Coder is: it not only writes, it writes correctly and cleverly.
Whether it is dynamic animation, 3D components, or interactive debugging and code fixes, it can quickly understand instructions and generate web code with clear structure and stable effects. Even more rare is that it can automatically fill in logic and interaction details based on context.
Although there is not yet a complete Agent framework, from the closed loop of "generation → execution → adjustment," it is already very close to an Agent-style workflow.
03. Some Sharing
After trying out Qwen3's three new models, the first thought that popped into my head was: "Qwen is not pursuing faster, but pursuing more accurate."
It is not simply stacking parameters, nor showing off with benchmark scores, but truly optimizing around real tasks and user needs.
In recent years, the large model track has been extraordinarily lively, with parameters getting larger, names getting more numerous, and update frequency getting faster, but doing things quickly does not mean doing them steadily.
The model is only the surface; behind it are chips, data, toolchains, and the entire closed loop from algorithms to products.
To continuously and stably produce high-quality results in complex systems, the one who truly goes far is often not the one with the most explosive highlights, but the one with the deepest investment and the strongest foundation.
Some companies can occasionally play a good card, but to maintain a stable rhythm and continuous output, they rely on long-term investment and systematic capability.
In recent years, Alibaba's pace of investment in the two main lines of AI and cloud computing has never stopped. Behind the frequent model updates, it is not only the algorithm team making efforts, but also long-term cultivation in infrastructure, computing platforms, engineering systems, and other aspects.
FromDAMO Academy's research, to the evolution of the Tongyi large model, and then to the foundational capabilities supported by Alibaba Cloud, Alibaba is indeed one of the few domestic companies capable of connecting the entire AI chain from algorithms to computing power and from models to products.
In the past we asked, can domestic AI catch up with GPT? Starting today, we can ask: why shouldn't we define industry standards?
Original link:Alibaba has just released the strongest reasoning model Qwen3-2507, fully evolved and comparable to Gemini 2.5 Pro