Alibaba releases its most powerful inference model, Qwen3-2507, comparable to the Gemini 2.5 Pro.

That's incredible! Is this the power of Alibaba?

Three days, three major releasesLarge ModelThey went online one after another and played a set of...AIThe industry's most hardcore "three-in-one technology release".

On July 22, the Alibaba Tongyi Qianwen team released a new version, Qwen3-235B-A22B-Instruct-2507, and released another version the following day.open sourceNew codeLarge Model Qwen3-Coder, tonight Qwen3-235B-A22B-Thinking-2507 It also made its grand finale appearance.

Within a week, fromGeneralintelligent, to codeintelligentFrom complex reasoning to advanced logic, the three core capabilities are fully covered.It just goes to show how comprehensive Alibaba is, with such a wide range of overseas operations.AIThe number of players is very small.

Not only the entire lineopen sourceIt also topped the charts in multiple international evaluations, directly competing with and even surpassing [previous rankings]. Gemini,Claude These international closed sourcesLarge Model.

Even Hugging Face CEO Clement publicly praised it.

However, I always trust my own user experience more; I'll know if it's awesome or not after testing it.

Six months ago,Qwen The exploration of a "hybrid thinking" model architecture, integrating speed and accuracy into a single model, has begun. This is a domestically developed...Large ModelA forward-looking attempt at "mindset switching".

This week, Qwen3 underwent a true "capability split": bidding farewell to hybrid thinking, it was directly split into two independent versions.

Although twoLarge ModelBoth models use the same 235B parameter architecture and the same 22B activation parameter, but their styles are completely different:

Qwen3-235B-A22B-Thinking-2507 focuses on in-depth 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 the performance of each model is more pure.

I've selected a few 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.

You can start using it right away.

Case 1: Informal Logic (Trap Question)

Prompt wordsIf I'm late, my salary will be docked.

I wasn't late today. So I won't get my pay docked, right?

Qwen3-235B-A22B-Thinking-2507 demonstrates the complete thought process.

Then they explained it step by step, so that I could understand what was going on.

The ending is a clear and concise summary.

The entire presentation allowed me to clearly see step-by-step how the model thinks and organizes its logic. To further clarify my understanding, it also provided real-world scenarios where wages might be deducted, such as leaving early, missing a clock-in, or making a work mistake.

Not only do you need to tell me the conclusion, but you also need to explain why you arrived at that conclusion.

Case 2: Risk and Return Choice

Prompt wordsYou have two choices:

A. 100% chance of receiving 80 yuan

B. There is an 80% chance of winning 100 yuan, and a 20% chance of winning 0 yuan.

Which one would you choose? Why?

Qwen3-235B-A22B-Thinking-2507 not only calculated the expected value correctly, but also has a good understanding of economics, accurately judging that option B might have a higher return, but also a higher degree of uncertainty, so he chose option A.

One interesting point is that Qwen3-235B-A22B-Thinking-2507 said, "As..."AI"I have no financial needs, but I can simulate a standard rational agent," which gives me a strong sense of "thinking role."

Case 3: Mental and Logical Testing Questions

Prompt wordsIdentifying pseudoscientific statements

"A certain brand claims that 'its drinking water can increase DNA frequency and enhance immunity.' How do you evaluate this kind of statement?"

This question seems like a matter of common scientific knowledge, but it's not actually...SimpleSince the model is designed to refute pseudoscience, judging from the answer, Qwen3-235B-A22B-Thinking-2507 not only saw through it immediately, but also gave me a quick science lesson.

Overall,Qwen3-235B-A22B-Thinking-2507 is not only "capable of thinking," but also possesses clear logic, rigorous reasoning, and a deep understanding of "the art of language."

Next, we'll test Qwen3-235B-A22B-Instruct-2507.

In QwenChat (https://chat.qwen.aiIn the ), select the Qwen3-235B-A22B-2507 model.

Case 4 Logical Reasoning

Prompt wordsGiven that "all birds can fly" is incorrect, can we conclude that "some birds cannot fly"? Please explain your reasoning using a syllogism.

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, according to my requirements, it uses a syllogism to deduce that "'not all birds can fly' is equivalent to 'some birds cannot fly'".

Does saying "not all S are P" necessarily mean "some S are not P"?

We first see a concise and clear answer: in classical logic this is equivalent, but in some non-classical logic or in specific contexts (such as the empty class problem), it may not hold true.

Being able to do this shows that Qwen3-235B-A22B-Instruct-2507 is not just memorizing logic rules, but is truly "understanding" the reasoning process.

Overall, Qwen3-235B-A22B-Instruct-2507 performs very well.Fast response, stable output, and clear logic.

Qwen3-Coder quickly gained popularity upon its release. Even Twitter founder Jack gave it a strong thumbs up, prompting a retweet response from the official Grok account.

I've chosen some commonly used scenarios, such as creating dynamic web pages, 3D cards, teaching demonstrations, and bug fixing, to see how Qwen3-Coder performs.

Case 1 Loading Animation

Prompt wordsCreate a full-screen loading animation component using HTML + CSS + JS. When loading, display a rotating planet or geometric shape animation, and fade into the main content after loading is complete.

Case 2: Parallax Effect for Webpage Scrolling

Prompt wordsWrite a webpage paragraph with a parallax scrolling effect, where the background image moves slowly as you scroll and the foreground text slides in. It would be suitable as a cover introduction section.

Case 3: Particle Background Animation

Prompt wordsCreate a cool particle background animation component using HTML + CSS + JS, where the particles will move according to the mouse movement direction.automaticIt flows and connects into a network, with a transparent background, and can be overlaid on any webpage.

Case 4 Animation Demonstration

Prompt wordsGenerate a dynamic animated webpage using multiple rotating circles (Fourier seriesA square wave is gradually constructed. Circles are connected end to end, with the end trailing out a dynamic trajectory to form a square wave. The background of the image is black, the lines are bright 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 Cards

Prompt wordsGenerate a 3D flip card display component, implemented using HTML and CSS. When the mouse hovers over the card, it flips back and forth to display different content, with smooth 3D animation.

Case 6: Fixing Bugs

Prompt wordsYou 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, it feels like there's no animation at all." Please find the cause, explain the problem, and fix the code (only modify 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>

This is how the page is displayed.

It quickly identified the specific cause and the solution, and eventually told me what adjustments it made.

The animation displays correctly, and the star trail effect appears!

After testing these cases, my biggest impression of Qwen3-Coder is that it not only writes code, but it also writes it correctly and skillfully.

Whether it's dynamic animation, 3D components, interactive debugging, or code fixing, it can handle it all.fastIt can understand instructions and generate well-structured, stable web page code. Even more remarkably, it can also adapt to context.automaticFill in the logic and interaction details.

Although it is not yet complete Agent The framework is there, but from the perspective of the closed loop of "generation → execution → adjustment," it is already very close. Agent This is a workflow.

After trying out the three new models of Qwen3, my first thought was: "Qwen is not pursuing faster speeds, but more accurate speeds."

noSimpleStacking parameters isn't about showing off benchmark scores; it's about genuinely optimizing for real-world tasks and user needs.

In recent years,Large ModelThe track is bustling with activity, with increasingly larger parameters, more and more names, and faster and faster updates, but doing things quickly does not mean doing them steadily.

The model is just the surface; behind it lies chips, data, toolchains, and the entire closed loop from algorithm to product.

To consistently and stably produce high-quality results in complex systems, those who truly go the distance are often not the ones with the most explosive breakthroughs, but rather the ones who invest the most deeply and build the strongest foundation.

Some companies may occasionally make a good move, but to maintain a stable pace and continuous output, they rely on long-term investment and systemic capabilities.

In recent years, Alibaba has AI Investment in both cloud computing and other related areas has remained constant. Behind the frequent model updates lies not only the efforts of the algorithm team, but also long-term dedication to infrastructure, computing platforms, and engineering systems.

fromDAMO AcademyResearch, to general meaningLarge ModelFrom its evolution to the foundational capabilities supported by Alibaba Cloud, Alibaba is indeed one of the few companies in China that can seamlessly integrate algorithms and computing power, and from models to products. AI One of the full-chain enterprises.

In the past, we asked, "Domestic products..." AI Can we catch up? GPTStarting today, we can ask: Why can't we define the industry standards?