Alibaba Launches New AI Model That Wins OpenAI and Google

Introduction to QWEN3-Thinking-2507: Artificial Intelligence Model That Changes New Standards

The renowned tech company, Alibaba, has launched the latest artificial intelligence (AI) model called QWEN3-235B-A22B-Thinking-2507. This model is part of the QWEN3-Thinking series and was developed with the aim of providing better performance than similar AI models. With the capabilities that are considered extraordinary, this model offers a variety of excellent features that make it the main choice for users and developers.

Excellent ability in various aspects

QWEN3-Thinking-2507 is designed to deal with a variety of complex tasks, including advanced mathematical and logic reasoning. The test results on the AIME25 benchmark show that this model is able to achieve a score of 92.3, which is much higher than other models such as Gemini-2.5 Pro from Google with a score of 88.0. This proves that this model has excellent analytical skills.

In addition, this model also shows the advantages in making code or coding. On the Livecodebench V6 benchmark, QWEN3-Thinking-2507 reached a score of 74.1, which surpassed the Gemini-2.5 Pro with a score of 72.5 and O4-Mini Openai with a score of 71.8. This shows that this model is very suitable for use in software development.

advantages in providing answers according to human preferences

This AI model is also considered capable of providing answers that are in accordance with human preferences. On the Arena-Hard V2 benchmark platform, QWEN3-Thinking-2507 scored 79.7, which is the highest number compared to other models such as Gemini 2.5 Pro with a score of 72.5 and Deepseek-R1-0528 with a score of 72.2. This shows that this model is very good at understanding and answering questions from a human point of view.

Mixture-Of-Experts Architecture (MOE)

Technically, QWEN3-Thinking-2507 is based on the Mixture-Of-Experts (MOE) architecture. This architecture is designed to be more efficient and flexible in handling various tasks and questions. With MOE, this model only activates a subset of 22 billion parameters, although it has a total of 235 billion parameters, for certain tasks. This allows the model to work optimally without overloading resources.

Extensive context window

This model also offers a fairly large context window, which is 262,144 tokens. This context window shows the number of tokens that can be processed in one interaction, including input and output tokens. With a wide context window, this model is capable of processing longer and more complex information.

new approach to development

Alibaba uses a new approach in developing QWEN3-Thinking-2507. The company no longer uses the “hybrid thinking” technique as in the previous QWEN3 model. This technique previously required developers to manually switch between fast instruction modes or deep reasoning (thinking), which were rated less efficient and consistent.

From community input, Alibaba decided to leave the Hybrid Thinking mode and train the Instruct and Thinking model separately. This approach is considered more optimal because the instruct model can be set for optimal speed and command execution, while the thinking model is trained for complex multi-step reasoning tasks.

access and price

QWEN3-Thinking-2507 is already available on the AI Hugging Face platform and can be accessed via the API. The price is set at 0.70 US dollars (around Rp. 11,453) per million input tokens and US$8.40 (around Rp. 137,443) per million output tokens. With competitive prices, this model is an attractive option for users and developers.