Qwen for
multilingual systems.

Most production AI systems end up wanting more than one model. A small one for classification and routing, a mid-size one for the bulk of the work, a large one for the hard minority of requests. Assembling that from three different families means three tokenisers, three prompt idioms, three evaluation baselines and three sets of quirks to learn.

Qwen's range lets you build that ladder inside one family. Prompts transfer with far less rework, an evaluation harness written for one size largely holds for another, and when you move a request class from the mid-size model to the small one — which is where a great deal of cost saving actually comes from — you are changing a parameter rather than rewriting an integration.

The multilingual coverage is the second reason we reach for it. If your corpus contains Chinese, Japanese or Korean alongside European languages, this family has, in our evaluations, handled the mix better than Western-trained alternatives of comparable size, and the difference shows up immediately on a fair evaluation set rather than needing a subtle one.

Qwen is developed by Alibaba, and the assessment is structurally the same one we set out on the DeepSeek page: separate the transfer question from the behaviour question from the artefact question. Self-hosting closes the first completely — the weights are a file and they transmit nothing. The second is testable on your own material and is workload-dependent. The third is only partially answerable, for every model from every jurisdiction.

Where an organisation has a standing policy against models of Chinese origin, that policy is the answer and we build with Mistral or Llama, which occupy adjacent ground. It is a governance decision owned by people who own the risk, and it is not our place to argue anyone out of it.

Where the deployment is a single size on a single class of hardware, the size ladder stops being an advantage and the decision comes down to capability and tooling on that one rung — which is frequently where Llama's deeper serving ecosystem wins.

Where the workload is at the hardest end of agentic or reasoning work, the leading proprietary models still lead, and a hybrid arrangement that routes that minority out is usually better than choosing one family for everything. See hybrid model routing.