It is QWEN: QWEN3-23B-A22B-ahinking-2507 Tops Tops

If the artificial intelligence industry has the equivalent of the movie “Summer Song” in the recordings industry – a success that holds in the warmer months here in the northern hemisphere and hears its role everywhere – the clear pasta of this title will go to the QWEN team in Ali Baba.

During the past week, the Frontier Model Ai Research Department issued a Chinese e -commerce no oneno twono threebut Four (!!) New open source models of artificial intelligence that offer records definition criteria, and are best even some leading royal options.

Yesterday night, QWEN team was crowned with a version QWEN3-235b-A22B-Theokeent-25507LLM model, which takes longer to respond to the unbroken LLM or “guidance”, participate in “idea chains” or self -reflection and self -realization that we hope will lead to more healthy and more comprehensive responses to the most difficult tasks.

In fact, the new QWen3-thethink-2507, where we will call it a short or now or now or follow the models with higher performance across many major criteria.

As an AI influence and the news complex Andrew Coran wrote on x: “The strongest thinking model in QWEN has arrived, which is at the limits.”

in Aime25 Standard-designer to assess the ability to solve problems in mathematical and logical contexts- QWEN3-ahinking-2507 With a degree 92.3OPENAI’s OPNA’s Sure exceeded (92.7) And Gemini-2.5 Pro (88.0).

The model also displays a leadership performance LiveCOOOOOBENCH V6and 74.1 registration, before Google Gemini-2.5 Pro (72.5), Openai O4-MINI (71.8)The performance of its previous version, which was published, greatly excelled 55.7.

in GPQAThe Standard of Multi -Options at the level of graduate studies, achieves the form 81.1Almost Deepseek-R1-0528 (81.0The mark of Gemini -2.5 of the professionals exceeds the professionals 86.4.

on Arena-Hard V2Which holds alignment and self-preference through winning rate 79.7And put it before all competitors.

The results show that this model does not exceed its predecessor in each major category, but also sets a new standard for what the open models of the sources that focus on thinking can achieve.

Far from “hybrid thinking”

The QWEN3-Hinking-2507 version reflects a broader strategic transformation by the QWEN team in Alibaba: Stay away from hybrid thinking models that ask users to manually switch between the conditions of “thinking” and “other than thinking”.

Instead, the team is now training separate models for thinking and education tasks. This separation allows to improve each model for its intended purpose – which indicates improving consistency, clarity and standard performance. The new QWEN3 model embodies the fully thinking of this design philosophy.

Besides, Launch QWEN QWEN3-Coder-480B-A35B-InstructThe parameter 480B model is designed for the functioning of complex coding work. It supports 1 million windows a symbolic context and outperforms GPT-4.1 and Gemini 2.5 Pro on Swe-Bused.

also Declare QWEN3-MTMulti -language translation model trained on trillions of symbols across more than 92 languages. It supports field adaptation, terminology control, and inference from only $ 0.50 per million symbols.

Earlier in the week, Establish the team QWEN3-235b-A22B-Instruct-25507It is an unique model that exceeds the Claud OPUS 4 on several criteria and provided a lightweight FP8 variable for a more efficient inference on restricted devices.

All models are licensed under APache 2.0 and are available through Hugging Face, Modelscope and QWEN API.

License: Apache 2.0 and the Foundation’s advantage for it

QWEN3-235b-A22B-Theacking-25507 is released below Apache 2.0 licenseA very tolerant license and a commercial friend that allows institutions to download, modify, self -control, control and integrate the model into ownership systems without restriction.

This contradicts only royal models or open versions of research only, which often require access to the application programming interface, imposing limits of use, or banning commercial publishing. For conscious organizations and teams that look forward to control of cost, cumin and data privacy, APache 2.0 license allows full flexibility and ownership.

Availability and pricing

QWEN3-235B-A22B-THEACKING-25507 is now available for free download on Embroidery and Models.

For those institutions that do not want or do not have resources and the ability to host the inferiorization model on their own devices or virtual cloud through API, VLLM and Sglang.

• Input price: 0.70 dollars per million symbols
• Output price: 8.40 dollars per million icons
• Free layer: 1 million symbols, valid for 180 days

The model is compatible with the frameworks worked QWEN-AgentAnd advanced publishing via OpenAi applications compatible with applications.

It can also be operated locally using transformers or merged into DEV chimneys through Node.js, Cli tools or organized claim facades.

Sampling settings include the best performance Temperature = 0.6and Top_p = 0.95And The length of the maximum product is 81,920 symbols For complex tasks.

Foundation applications and future expectations

Through its strong standard performance, the ability of long context, and permissible licensing, QWEN3-THINKING-2507 is especially suitable for use in the AI systems for the institution that includes thinking, planning and decision support.

The broader ecosystem QWEN3 – including coding, education and translation models – extends to the call to technical groups and business units that look forward to integrating artificial intelligence through vertical such as engineering, Emiratisation, customer support and research.

QWEN team’s decision to issue specialized models for distinguished use cases, supported by technical transparency and community support, indicates a deliberate shift towards construction Open infrastructure, performance and preparation for production.

Since more institutions are looking for alternatives to black boxes of API, the QWEN series of Alibaba increases itself as an open-ended source of smart systems-with a large scale control and ability.