New Qwen 3.5 open source models released; Suitable for running on personal systems

Alibaba’s artificial intelligence development team introduced the new Qwen 3.5 series of language models, which brings the features of advanced and expensive models to PCs for free and locally.

The Qwen 3.5 series is a collection of four major language models that offer equivalent or better performance than their paid competitors, and the developers have released three versions of them as fully open source. Alibaba has licensed Qwen 3.5 with 35 billion parameters, 122 billion parameters, and 27 billion parameters under the Apache 2.0 license for free commercial use on HoggingFace and ModelScope platforms. The Qwen3.5-Flash model is also available only as an API at a much lower price than the US models.

These free models have been able to beat powerful models like GPT-5 Mini and Cloud Sonnet 4.5 easily in independent benchmarks.

Qwen 3.5 models suitable for running on personal systems

Alibaba engineers have designed these models based on the advanced architecture of combined specialists to provide users with the highest speed and accuracy with the least hardware involvement. For example, the 35 billion parameter flagship version only activates 3 billion parameters for word processing at any given moment.

This architecture uses 256 different specialists who divide the processing among themselves. In addition, the development team has compressed these models in such a way that even with a drastic reduction in the amount of information, their accuracy experiences almost no loss and the quality of the output is maintained.

You need relatively powerful systems to run Qwen 3.5 large models; For example, to run the most powerful open source version of this artificial intelligence, you must have a graphics card with 32 GB of VRAM. Although not everyone may have such a graphics card, it is still impressive to run such a powerful model with such a system. This model can process a 1 million token content window (reading and writing long files) and previously you would have had to have gigantic servers to run it.

Alibaba’s unique compression technology enables independent developers to process large data sets on home systems without the need for expensive server infrastructure. Also, the 27 billion parameter model of this family is optimized for higher efficiency and can easily manage 800,000 tokens.