Ali Tongyi Thousand Questions released the largest model to date - Qwen3-Max-Preview, the number of parameters over 1 trillion

On September 5, Ali went liveQwenThe strongest model of the 3 seriesQwen3-Preview version of Max, which is Ali's largest model to date, with over 1 trillion participants. The model is now available on the Ali Bailian platform and is available for free on the Tongyi Thousand Questions app and Qwen Chat.

According to the Hundred Refined Platform.Qwen3-Max-PreviewCompared to the 2.5 series, the overall generalization ability has been substantially improved, with significant enhancements in Chinese and English general text comprehension, complex instruction following, subjective open tasking, multilingualism, and tool invocation; and fewer model knowledge illusions.

Just yesterday, the official Qwen X account teased the upcoming release of one of the most powerful and intelligent members of the Qwen3 family. Today, this model went live and its reviews were released.

Qwen3-Max-Preview is said to have outperformed Claude-Opus 4 in the General Knowledge (SuperGPQA), Mathematical Reasoning (AIME25), Programming (LiveCodeBench v6), Human Preference Alignment (Arena-Hard v2), and Comprehensive Competency Assessments (LiveBench) rubrics ( Non-Thinking), as well as Kimi-K2, DeepSeek-V3.1 and Ali's previous open source best Qwen3-235B-A22B-Instruct-2507.

The presentation of Qwen3-Max on OpenRoute, an AI model aggregation platform, mentions that it offers significant improvements in inference, command execution, multi-language support, and long-tail knowledge coverage; as well as higher accuracy in mathematical, programming, logic, and scientific tasks. The model supports over 100 languages, has stronger translation and common sense reasoning, and is optimized for retrieval-enhanced generation (RAG) and tool invocation, but does not include a dedicated "think" mode.

Wisdom things the first time on the Tongyi Qianqian web end of the Qwen3-Max-Preview experience, found that the model in the text comprehension, as well as mathematical and programming ability is excellent, and the response speed is very fast.

First let Qwen3-Max-Preview generate a ball collision simulator, we enter the prompt word: "A circle with two balls inside, a black one and a white one, the white ball follows the position of the free fall, and bounces when it touches the boundary, and at the same time generates a randomly positioned white ball, the black ball bounces when it touches the boundary, and the white ball bounces when it touches the boundary. ball will get a little bigger, please simulate it."

Only Qwen3-Max-Preview quickly outputs this program, simulating the movement of the two types of balls, with the black ball eventually expanding to the point where it engulfs the white ball.

When we ramped up the difficulty to allow Qwen3-Max-Preview to perform a strength and speed population simulation, and continued to optimize this simulator, we found that Qwen3-Max-Preview was able to achieve a fast and accurate simulation, able to do in a few seconds what it might take a full-fledged programmer most of a day to do.

We enter the prompt, "There are two populations, population a focuses on the development of strength and population b focuses on the development of speed, model the interaction between the two populations and give a description."

As you can see below, Qwen3-Max-Preview understood what I meant and gave a more accurate simulation even though I gave very vague cues.

In the above simulation, I realized that the speed-oriented populations were being killed off too quickly, and I further wanted them to be able to "run away". I typed in the prompt: "The speed-oriented populations were killed off too quickly, and each of them should have some ability to avoid danger."

Qwen3-Max-Preview then outputs the following "Strength and Speed Population Simulation (Enhanced)", which accurately simulates the situation of "no one can kill anyone" with the ability to avoid dangerous balls.

If you can only run away and not counterattack, sooner or later you will still be taken out. So I asked for a collaborative offense ability for speedy populations, and entered the prompt word: "When speedy populations are united, they can can take out single power individuals, please add this ability and simulate it again."

Qwen3-Max-Preview still works well, outputting a "Power and Speed Population Simulation (Collaboration Version)", which simulates the ability of the little green ball to fend off the red ball when it has the ability to collaborate, but the two sides are still very much at a stalemate.

As the simulation progressed, the populations on both sides got smaller and smaller, so we further asked Qwen3-Max-Preview to give them the ability to reproduce by typing in the cue word: "When both of them take out each other's individuals, they can build up nutrients, reproduce themselves, and continue the simulation."

So, Qwen3-Max-Preview outputs a "Power and Speed Population Simulation (Resource and Reproduction Version)", from which it can be seen that both types of balls start to fission on their own, in which case the red balls can no longer outcompete the green ones.

So I typed again:

"Strength populations have been found to be so weak they can't catch the opposite side, please give them the ability to work as a team as well and be able to round up speedsters."

Qwen3-Max-Preview outputs a "Power and Speed Population Simulation (Bi-directional Collaboration Version)", in which the small green ball and the small red ball form a huddle, which creates a "huddle" on both sides.

Through this interesting little experiment we found that Qwen3-Max-Preview was able to successfully understand the user's intent even when the cue words were very ambiguous in their meaning.

In particular, expressions such as "avoid danger", "unite", "collaborate", and "reproduce" are relatively abstract, and the corresponding practical meanings are complex. The actual meanings of these expressions are very complex, and their implementation involves the adjustment of many parameters. However, Qwen3-Max-Preview accurately understands the semantics and the logic behind them within a few seconds, and completes the programming of the simulation experiments, demonstrating its excellent ability in complex reasoning, instruction execution, mathematics, programming, and other abilities.

As shown by the Hundred Refinement platform, in terms of pricing, Qwen3-Max-Preview supports 256k contexts and takes step billing based on the number of tokens entered:

Input 0-32k token price: $0.006/thousand token input, $0.024/thousand token output.

Input 32k-128k token price: $0.01/thousand token input, $0.04/thousand output.

Input 128k-252k token price: $0.015/thousand token input, $0.06/token output.

Compared to Qwen-Max-0919's price of $0.02/thousand token input and $0.06/thousand token output, Qwen3-Max-Preview's pricing is more hierarchical, with higher performance but at a more affordable price.

Experience Address:https://chat.qwen.ai

AliCloud Hundred Refined API Service:https://bailian.console.aliyun.com/?tab=model#/model-market

Conclusion: the oversized Qwen3 model, demonstrating the effect of scaling up

The breakthrough in the model layer is becoming the first trump card of Ali's AI transformation. In internal tests and early user evaluations, Qwen3-Max-Preview has demonstrated broader knowledge, better dialog capabilities, and stronger performance in terms of Agent tasks and instruction following.

lit. ten thousand questions on general principles (idiom); fig. a long list of questions and answersLarge model open sourceClosed-source two-handedly, has represented the new height of China's large model technology. qwen3-Max-Preview refreshes the aliLarge model parametersThe new record, with its attempt to prove the effect of scaling up with even more robust performance - bigger models have more performance.

Article source: Wisdom