Alibaba launches Qwen 3.5, a commitment to multimodal agents for reduced budgets

Chinese-origin artificial intelligence technology is increasingly gaining traction in a sector where, in theory, North American companies such as OpenAI, Anthropic, Google, or Microsoft were expected to lead. We have witnessed this just a few days ago, with the enormous impact generated by videos created by Seedance 2.0 (owned by Bytedance, parent company of TikTok), which have even prompted legal threats from companies such as Disney or Paramount, who view their intellectual property as being threatened to some extent.

Now, another major and well-established player in the Chinese business landscape, Alibaba, is demonstrating that it does not intend to lag behind in this rapidly accelerating arms race for control of AI. Qwen 3.5 is here.

Qwen 3.5: a commitment to multimodality

This is not the first time we have discussed Qwen, the family of artificial intelligence models owned by Alibaba, which has been advancing a highly demanded technology for several years. In this instance, the company has just introduced Qwen3.5, its new cutting-edge model series. The launch begins with the open-sourcing of Qwen3.5-397B-A17B (known as Qwen3.5-Plus), a fundamentally multimodal model.

According to the company’s official statement, thanks to its ability to interpret text, images, and video simultaneously, the model can function as a visual agent capable of interacting with smartphone and computer interfaces. Furthermore, its analytical capabilities extend to long-form videos of up to two hours, thus opening new avenues for content analysis and the resolution of complex scientific problems.

Another of its primary features is its orientation toward programming, automating complex workflows and, for example, instantly converting simple interface sketches into functional code through Qwen Code. Additionally, it enables the creation of websites more easily by converting simple text instructions into operational code (as you can observe in the video below): “With Qwen3.5 as the underlying model, Qwen Code supports ‘vibration coding’, converting natural language instructions into code, iterating projects in real time, and managing creative tasks such as video generation or other resources. Together, Qwen Code and Qwen3.5 provide an optimized experience for both daily and exploratory programming.”

Alibaba assures that Qwen3.5 can be used in conjunction with today’s viral model, OpenClaw, to optimize coding tasks. By integrating with OpenClaw as an external agent environment, Qwen3.5 can perform web searches, collect information, and generate structured reports, combining its reasoning and tool use capabilities with the OpenClaw interface for a seamless coding and research experience.

Coding: Website pic.twitter.com/ILL5PvEEKb

— Qwen (@Alibaba_Qwen) February 16, 2026

Finally, linguistic versatility is presented as another pillar of this version. The model supports 201 languages and dialects, nearly doubling its predecessor’s capability.

Agentic capabilities for modest budgets

Alibaba states that the engineering behind this new family of models enables the Qwen3.5-397B-A17B to match the performance of much larger models—such as the Qwen3-Max, which exceeds one trillion parameters—but with significantly greater efficiency. In practical terms, this allows developers to scale multimodal applications without budgets growing proportionally. As explained by the company: “The launch advances Alibaba’s drive toward agentic AI: building models that are not only capable, but also efficient enough to scale broadly, allowing developers and enterprises to deploy multimodal applications without proportionally increasing computational budgets.”

The relevance of this technology is already tangible at events on a global scale. Recently, the International Olympic Committee presented its first language model-based system—developed on Qwen—at the Milano Cortina 2026 Winter Olympics, designed to enhance the fan experience and optimize the logistics of the Olympic ecosystem. Currently, the model is available to the global community via platforms such as Hugging Face, GitHub, and Alibaba Cloud.

Developers can access Qwen3.5-397B-A17B through Hugging Face, GitHub, and ModelScope, and experience it via Qwen Chat. The model is also available through API in the Alibaba Cloud model development platform, Model Studio, and additional Qwen3.5 models are expected to be released as open source in the coming weeks.

Image: Gemini