Want a smart insight into your inbox? Sign up for our weekly newsletters to get the only thing that is important to enterprise AI, data, and security leaders. Subscribe now
If the recording industry’s “summer song” in the AI industry is equivalent to the “summer song” – this is a hit that is found here in the northern hemisphere in the hot months and is heard everywhere – the clear honor for this title will go to Alibaba’s coin team.
Only during the last week, the Frontier Model AI Research Division of Chinese e -commerce has not released one, no, no, no, not ThreeBut Four (!!) The new open source Generative AI model that offers a record setting benchmark, and even improve some well -known proprietary powers.
Last night, The Kevin team contested it with the release QWEN3-235b-A22b-Thenking-25507This is the reasoning of the latest language model (LLM), which takes more time to engage in “chains thinking” or self -reflection and self -checking or responding to “guidance” LLM, which in turn hopes that more difficult and comprehensive reactions will cause more difficult tasks.
In fact, the new Qwen3 thinking -25507, as we call it briefly, now guide or trails highly performing models in several large benchmarks.
AI Impact Series returning to San Francisco – August 5
The next step of the AI is here – are you ready? Block, GSK, and SAP leaders include for a special look on how autonomous agents are changing enterprise workflows-from real time decision-making to end to automation.
Now secure your place – space is limited:
As AI Inflox and News Agriculture Andrew Corne wrote on X: “Kevin’s strongest reasoning model has arrived, and it is on the border.”
I Aime25 Benchmark-state and logical context are designed to assess the ability to solve the problem. QWEN3-Inking-25507 leads to all reported models With a score 92.3Reducing both the o4-mini of Openi (92.7) And Gemini -2.5 Pro (88.0,
The model also displays commanding performances Livecodebench v6For, for, for,. Google Gemini -2.5 Pro (72.5) beyond 74.1 scoring, Open AO 4 -Meni (71.8)And significantly improve its first version, which posted 55.7.
I GPQAGraduate Surface receives a quality, model for multiple selected questions 81.1Matching almost depapsek-r1-0528 (81.0) And the top sign of the gym -2.5 Pro 86.4.
On Arena Hard V2Which reviews alignment and saplus preference through winning rates, QWEN3-Inking-25507 scores 79.7Keep it ahead of all rivals.
The results show that this model not only crosses its predecessor in every major category, but also sets a new standard to achieve an open source, reasoning model.
A shift away from ‘hybrid reasoning’
The release of QWen3-Thenking-25507 reflects a wider strategic change by Alibaba’s Kevin team: away from hybrid reasoning models, consumers need to be togelled between “thinking” and “non-thinking” methods.
Instead, the team is now training separate models for reasoning and instructions. This separation allows every model to improve for its own purpose. The new Kevin 3 -thinking model completely embraces the philosophy of this design.
As well, Kevin launched QWEN3-CODER-480B-A35B-InstructA 480b-parameter model made for complex workflows of coding. It supports 1 million token context Windows and performs GPT -4.1 and Gemini 2.5 Pro on SWE Bench certified.
Plus Was announced Qwen3-mtA multi -linguistic translation model trained on trillions of tokens in 92+ languages. This domain supports adaptation, terms control, and only $ 50 per million tokens per million.
At the beginning of the week, the team released QWEN3-235b-A22b-Instruct-25507An irrational model that surpassed Claude Ops 4 on several benchmarks and introduced a lightweight FP8 variety for the compulsive hardware to be more efficient.
All models are licensed under Apache 2.0 and are available via the throat faces, models cope, and Kevin API.
Licensing: Apache 2.0 and its enterprise advantage
Been released under the QWEN3-235b-A22b-Inking-25507 Apache 2.0 LicenseA highly legitimate and commercial -friendly license that facilitates businesses to download, modify, edit, host, good tone and merger in proprietary systems without any restriction.
It stands unlike proprietary models or open releases with just research, which often requires access to API, imposes limits, or banned commercial deployment. For compliance organizations and teams who want to overcome cost, delay, and data privacy, Apache 2.0 licensing enables full flexibility and ownership.
Availability and pricing
QWEN3-235b-A22B-Thenking-25507 is now available for free download The hugs face And Models.
For businesses who do not want the resources and ability to identify the model on their hardware or virtual private cloud through Alibaba Cloud’s API, VLM, and Sglang.
- Input price: 70 0.70 per million tokens
- Production price: 40 8.40 per million tokens
- Free levels: 1 million tokens valid for 180 days
The model agent is compatible with the framework Qwen-addentAnd the open AI supports the latest deployment through APIS.
It can be operated locally using a transformer framework or node. The JS, CLI tools, or structural indicator interfaces can be integrated into DEV steaks.
Samples taking for excellent performance includes Temperature = 0.6For, for, for,. Top_P = 0.95And Maximum output length 81,920 token For complicated tasks.
Enterprise applications and the future outlook
With its strong benchmark performance, long context capability, and legitimate licensing, QWen3-insking-25507 is appropriate to use the use of planning, planning and decision-help in the Enterprise AI system, especially in the enterprise AI system.
Widely QWen3 Environmental System – which includes coding, instructions, and translation models – Forthar has extended appeal to technical teams and business units seeking AI in a vertical part such as engineering, localization, customer support, and research.
The decision to issue a special model for various issues of the Kevin team use, which supports technical transparency and community support, indicates deliberate change towards the building. AI infrastructure ready for open, performance, and production.
Since more businesses have sought alternatives to API Good, black boxes, Alibaba’s Kevin series rapidly positions itself as a viable open source foundation for the intelligent system.
