Alibaba's Open-Source Qwen AI Model Breaks Records in Reasoning
The Qwen team at Alibaba has unveiled a new version of their open-source reasoning AI model, showcasing remarkable benchmark results.
Introducing Qwen3-235B-A22B-Thinking-2507. For the last three months, the Qwen team has been intensively scaling up what they refer to as the model’s "thinking capability," striving to enhance both the quality and depth of its reasoning processes.
The outcome is a model that truly shines in the most demanding areas: logical reasoning, complex mathematics, scientific challenges, and advanced coding. In fields that typically demand human expertise, this latest Qwen model is now setting a new bar for open-source AI.
On reasoning benchmarks, Qwen's newest open-source AI model scores 92.3 on AIME25 and 74.1 on LiveCodeBench v6 for coding. It also performs strongly in broader capability evaluations, achieving a 79.7 on Arena-Hard v2, a metric that assesses alignment with human preferences.

Fundamentally, this is a large-scale reasoning AI model from the Qwen team, featuring a total of 235 billion parameters. However, it employs a Mixture-of-Experts (MoE) architecture, meaning only a subset of these parameters—approximately 22 billion—are active at any given time. Imagine it as a vast team of 128 specialists on standby, with only the top eight experts for a particular task actually working on it.
One of its standout attributes is its exceptional memory capacity. Qwen's open-source reasoning AI model natively supports a context length of 262,144 tokens, providing a significant advantage for tasks requiring the comprehension of extensive information.
For developers and enthusiasts, the Qwen team has streamlined the getting-started process. The model is accessible on Hugging Face and can be deployed using tools like sglang or vllm to set up a personal API endpoint. The team also highlights their Qwen-Agent framework as the optimal method for leveraging the model's tool-calling functionalities.
To achieve peak performance with this open-source AI reasoning model, the Qwen team offers several recommendations. They advise an output length of around 32,768 tokens for standard tasks, but for highly complex problems, increasing this to 81,920 tokens allows the AI sufficient space to "think." They also suggest using explicit instructions in your prompts, such as requesting a "step-by-step reasoning" approach for mathematical problems, to obtain the most precise and well-organized responses.
The launch of this new Qwen model delivers a powerful, open-source reasoning AI capable of competing with leading proprietary models, particularly in tackling intricate, intellectually demanding challenges. It will be fascinating to observe what the developer community creates with this technology.
See also: AI Action Plan: US leadership must be ‘unchallenged’
Interested in deepening your knowledge of AI and big data from industry experts? Attend the AI & Big Data Expo in Amsterdam, California, and London. This comprehensive event runs alongside other major conferences, including the Intelligent Automation Conference, BlockX, Digital Transformation Week, and the Cyber Security & Cloud Expo.
Discover more upcoming enterprise technology events and webinars powered by TechForge here.
Related article
Warner Music acquires AI attribution startup Sureel AI
Warner Music Group (WMG) confirmed on Wednesday that it is acquiring Sureel AI, an artificial intelligence attribution startup. Sureel’s proprietary technology generates “AI DNA” for musical tracks, deconstructing them into constituent elements to tr
Amazon introduces Alexa for Shopping while pushing Rufus to the background
Amazon has launched Alexa for Shopping, merging its Rufus shopping chatbot with Alexa+ across the app, website, and Echo Show devices.The assistant answers product queries, compares items, tracks prices, and supports shopping reminders. It also handl
Microsoft, Azure and AI Tech Combat California Wildfire Risks
Microsoft invests in AI-driven wildfire detection, with Juan Lavista Ferres, CVP and Chief Data Scientist, discussing strategies to mitigate environmental damage.According to NASA, climate change impacts everyone on Earth, manifesting as rising tempe
Related Special Topic Recommendations
Comments (1)
0/500
The Qwen team at Alibaba has unveiled a new version of their open-source reasoning AI model, showcasing remarkable benchmark results.
Introducing Qwen3-235B-A22B-Thinking-2507. For the last three months, the Qwen team has been intensively scaling up what they refer to as the model’s "thinking capability," striving to enhance both the quality and depth of its reasoning processes.
The outcome is a model that truly shines in the most demanding areas: logical reasoning, complex mathematics, scientific challenges, and advanced coding. In fields that typically demand human expertise, this latest Qwen model is now setting a new bar for open-source AI.
On reasoning benchmarks, Qwen's newest open-source AI model scores 92.3 on AIME25 and 74.1 on LiveCodeBench v6 for coding. It also performs strongly in broader capability evaluations, achieving a 79.7 on Arena-Hard v2, a metric that assesses alignment with human preferences.

Fundamentally, this is a large-scale reasoning AI model from the Qwen team, featuring a total of 235 billion parameters. However, it employs a Mixture-of-Experts (MoE) architecture, meaning only a subset of these parameters—approximately 22 billion—are active at any given time. Imagine it as a vast team of 128 specialists on standby, with only the top eight experts for a particular task actually working on it.
One of its standout attributes is its exceptional memory capacity. Qwen's open-source reasoning AI model natively supports a context length of 262,144 tokens, providing a significant advantage for tasks requiring the comprehension of extensive information.
For developers and enthusiasts, the Qwen team has streamlined the getting-started process. The model is accessible on Hugging Face and can be deployed using tools like sglang or vllm to set up a personal API endpoint. The team also highlights their Qwen-Agent framework as the optimal method for leveraging the model's tool-calling functionalities.
To achieve peak performance with this open-source AI reasoning model, the Qwen team offers several recommendations. They advise an output length of around 32,768 tokens for standard tasks, but for highly complex problems, increasing this to 81,920 tokens allows the AI sufficient space to "think." They also suggest using explicit instructions in your prompts, such as requesting a "step-by-step reasoning" approach for mathematical problems, to obtain the most precise and well-organized responses.
The launch of this new Qwen model delivers a powerful, open-source reasoning AI capable of competing with leading proprietary models, particularly in tackling intricate, intellectually demanding challenges. It will be fascinating to observe what the developer community creates with this technology.
See also: AI Action Plan: US leadership must be ‘unchallenged’
Interested in deepening your knowledge of AI and big data from industry experts? Attend the AI & Big Data Expo in Amsterdam, California, and London. This comprehensive event runs alongside other major conferences, including the Intelligent Automation Conference, BlockX, Digital Transformation Week, and the Cyber Security & Cloud Expo.
Discover more upcoming enterprise technology events and webinars powered by TechForge here.
Warner Music acquires AI attribution startup Sureel AI
Warner Music Group (WMG) confirmed on Wednesday that it is acquiring Sureel AI, an artificial intelligence attribution startup. Sureel’s proprietary technology generates “AI DNA” for musical tracks, deconstructing them into constituent elements to tr
Microsoft, Azure and AI Tech Combat California Wildfire Risks
Microsoft invests in AI-driven wildfire detection, with Juan Lavista Ferres, CVP and Chief Data Scientist, discussing strategies to mitigate environmental damage.According to NASA, climate change impacts everyone on Earth, manifesting as rising tempe





Home






