Alibaba releases Qwen 3.5 Small, compact models built to run directly on devices

For the past two years, the artificial intelligence industry has lived by a simple formula: the more parameters, the smarter the model. Alibaba has just offered an alternative. The Qwen team has released Qwen 3.5 Small, a series of compact language models, a family ranging from 0.8 to 9 billion parameters, designed specifically to run on users' devices without calling cloud servers.

To understand why this release matters, it helps to look at the context. For the past year and a half, the world's largest labs, including OpenAI, Google and Anthropic, have competed to build ever larger "frontier" models with hundreds of billions and even trillions of parameters. These models deliver impressive results, but they have a fundamental limitation: they require powerful server infrastructure and run exclusively through the cloud. Every user request travels to a remote data center and back, which creates latency, dependence on an internet connection and data privacy concerns. Alibaba decided to attack exactly this problem.

The philosophy behind Qwen 3.5 Small is summed up concisely: "More Intelligence, Less Compute." Behind this marketing slogan lies serious engineering work. The models in the family span a range from 0.8 billion parameters, which lets them run even on budget smartphones, to 9 billion, a size comfortable for modern laptops and tablets with enough RAM. The key question, which open benchmarks cannot yet answer conclusively, is how "smart" these models really are compared with competitors of similar size. However, previous generations of Qwen consistently delivered competitive results, and there is no reason to believe the new series will be an exception.

It is important to understand that Alibaba is not a pioneer here, but it may be the most ambitious player. Microsoft has already released its Phi series, Google has Gemma, and Meta* is developing compact versions of Llama. Qwen 3.5 Small stands out, however, for the breadth of its lineup: offering a whole family of models of different sizes under a single architecture gives developers flexibility of choice. The creator of a mobile note-taking app can take the 0.8-billion-parameter version for basic autocorrect, while a smart speaker maker can use a 4-billion model for a full voice assistant. The 9-billion-parameter model can handle tasks that required a cloud API call just a year ago.

The release also has a strategic dimension for Alibaba. The Chinese tech giant is fighting for global influence in AI infrastructure, and open models are one of its main tools in that fight. Every developer who integrates Qwen into an app becomes part of Alibaba's ecosystem. With US sanctions limiting Chinese companies' access to advanced chips for training giant models, the bet on efficient compact models looks not just like a technological choice but like a strategy that is both forced and farsighted. If you cannot get the best hardware to train trillion-parameter monsters, it makes sense to learn to squeeze the most out of smaller models.

The consequences for the industry could be significant. On-device models solve several painful problems at once. Privacy: user data never leaves the phone. Speed: no delay from a network request. Availability: the AI works without an internet connection. Cost: developers do not have to pay for every API call. For billions of users in developing countries, where mobile internet is unstable and expensive, local AI on a smartphone is not a luxury but the only realistic path to the technology.

Of course, compact models will not replace frontier ones. Complex reasoning, long coherent text generation and multimodal analysis all remain the territory of large models for now. But the vast majority of everyday tasks, such as summarization, translation, autocomplete, classification and simple dialogue, are well within reach of models with a few billion parameters. And these are the tasks that define the user experience for hundreds of millions of people.

The release of Qwen 3.5 Small confirms a trend that will shape the AI industry in the coming years: the future of artificial intelligence is not only in the cloud. It is in every user's pocket, in every device capable of running inference locally. Alibaba has placed its bet, and now the ball is in its competitors' court.

*Meta has been designated an extremist organization and is banned in Russia.