Alibaba Unveils Qwen-3.5, Escalating Global AI Model Competition

Alibaba Cloud has unveiled Qwen-3.5, its next-generation open artificial intelligence model, which features multimodal capabilities and open weights. This release signals Alibaba’s ambition to anchor the next phase of global AI deployment, sharpening the global AI race between China and the US.

Credit: QWEN.AI

The new 3.5-series models became available on Alibaba’s cloud platform Model Studio. One model, Qwen-3.5-Open-Source, contains 397 billion parameters and showed significant improvement over the company’s previous flagship model, Qwen-3-Max-Thinking, despite the latter’s larger size of over 1 trillion parameters.

According to self-reported benchmark scores, Qwen-3.5-Open-Source performed on par with leading models from OpenAI, Anthropic, and Google DeepMind. However, this comparison was not made with the three US heavyweights’ latest models.

A closed-source version, Qwen-3.5-Plus, achieved performance "on par with state-of-the-art leading models." It has a listed context window, the amount of data it can process at any given time, of 1 million tokens, one of the largest in the industry.

Parameters are mathematical variables that shape how an AI system learns and reasons. Alibaba Cloud, the AI and cloud computing arm of Alibaba Group Holding, built Qwen-3.5 with native multimodal capabilities for the first time, allowing it to understand text, images, audio, and video within a single system.

The model family also adopts the company’s latest architecture, first previewed through an experimental system called Qwen3-Next. This design aims to improve computational efficiency and reduce operating costs, helping the model achieve "a new benchmark for capability per unit of inference cost."

Model weights, which are parameters encoding the systems’ capabilities or "intelligence," were released on developer platforms Hugging Face and Alibaba’s own ModelScope. This enables users with sufficient hardware to download and run the models locally.

China has positioned itself as a leading force in open-source AI, differentiating its approach from the more closed-model strategies favoured by many Silicon Valley giants that do not publicly release model weights. Download data from Hugging Face showed Chinese open models overtook US counterparts in global adoption, with DeepSeek and Qwen accounting for the bulk of that growth.

Qwen has gained traction for its breadth of use cases, as Alibaba typically releases multiple versions of the same base models at different sizes to drive uptake across developers and enterprises. Its multilingual support has also contributed to its adoption.

The new Qwen-3.5 models added 82 new languages and dialects from the previous Qwen generation, supporting 201 languages and dialects, including niche languages such as Hawaiian and Fijian, according to the latest announcement.

Open-model expert Nathan Lambert noted that Qwen downloads on Hugging Face exceeded those of all other major open models combined. Such features have helped establish Qwen as a de facto “open model standard.”

China’s growing prominence in the open-model ecosystem prompted the Trump administration to prioritise the global diffusion of US-developed open models. However, AI and semiconductor analyst Lennart Heim suggested that real-world deployment is a more telling metric than downloads alone.

Heim added that the national-security implications of open Chinese models remain uncertain, even as some US start-ups and academic institutions deployed them locally. Notably, Alibaba does not open-source all of its systems.

Its largest models by parameter count, known as the Max series, remain closed and closely integrated with the company’s flagship consumer app, Qwen. Alibaba has also pursued an aggressive commercialisation strategy, positioning open models at the centre of its cloud business.

This strategy involves bundling data hosting and model-inference services for developers and enterprises worldwide. Lin Junyang, technical lead of Alibaba Cloud’s Qwen team, suggested on X that they would release more open-weight models.

---

• Alibaba Cloud launched Qwen-3.5, an open AI model with multimodal capabilities and open weights.

• Two versions, Qwen-3.5-Open-Source and Qwen-3.5-Plus, demonstrate strong performance against leading AI models.

• Model weights are openly released on platforms like Hugging Face and ModelScope, aligning with China's open-source AI strategy.

Source: SCMP