How do you try Alibaba's Qwen 2.5 models?

On September 19, 2024, Alibaba's Qwen team released Qwen 2.5, a family of language models you can download and run yourself. These are open-weight models: the trained numbers that make up the model (its weights) are published, so anyone can run it on their own hardware, with no account and no per-message fee. The family comes in seven sizes, from 0.5 billion to 72 billion parameters (the adjustable values a model learns in training; more usually means more capable, but also heavier).

Alongside the general models, the team released Qwen2.5-Coder for programming and Qwen2.5-Math for mathematics. The general models can read up to 128,000 tokens of context, write up to about 8,000 tokens in one reply, and are listed as supporting more than 29 languages, including English and French.

Pick a size your computer can handle

A model has to fit in your computer's memory to run at a usable speed. Ollama stores most models in a compressed form called quantization, which rounds the weights to fewer digits so files shrink a lot with a modest loss in quality. Here are the default download sizes in Ollama's library this month:

Ollama tag	Parameters	Download size	Good for
`qwen2.5:0.5b`	0.5 billion	398 MB	Quick tests, very small machines
`qwen2.5:1.5b`	1.5 billion	986 MB	Simple rewriting and sorting
`qwen2.5:3b`	3 billion	1.9 GB	Light everyday use on a laptop
`qwen2.5` (7b)	7.6 billion	4.7 GB	A sensible default for most recent computers
`qwen2.5:14b`	14 billion	9.0 GB	Better quality, needs a well-equipped machine
`qwen2.5:32b`	32 billion	20 GB	High-memory workstations
`qwen2.5:72b`	72 billion	47 GB	Servers or very large Macs

Your computer needs free memory beyond the file size itself, for the conversation and the program running it. If replies crawl along a word at a time, step down one size.

Run it with Ollama in three steps

Download and install Ollama from ollama.com.
Open a terminal (Terminal on a Mac, PowerShell on Windows) and type ollama run qwen2.5. The first run downloads the 7B model; later runs start in seconds.
Type a question at the prompt. Type /bye to leave.

To try another size, add its tag, for example ollama run qwen2.5:3b. Everything runs on your own machine, so your text never leaves it. That makes local models a good choice for drafts you would rather not send to an online service.

If you write code, the Qwen announcement shows the Hugging Face route as well. The model files live under names such as Qwen/Qwen2.5-7B-Instruct, and the Transformers library (Hugging Face's Python toolkit for running models) can load them in a few lines. "Instruct" means the version tuned to follow instructions and chat; the "base" versions are the raw pretrained models, meant for further training.

Check the licence before you build on it

Licences differ by size. All the open Qwen 2.5 models use the Apache 2.0 licence, a permissive licence that allows commercial use, except the 3B and 72B versions, which use Alibaba's own Qwen licence. Ollama's page repeats the same split. If you plan to use a model in a product or for a client, open the licence file in that size's Hugging Face repository and read it, or pick one of the Apache sizes such as 7B or 14B.

A worked bilingual test

A fair test uses the same tasks in both languages, on the kind of work you really do. Here is a short set you can paste one at a time into ollama run qwen2.5:

Prompt

1. Explain in three sentences what a heat pump does, for a homeowner with no technical background.
2. Explique en trois phrases ce que fait une thermopompe, pour un propriétaire sans connaissances techniques.
3. Translate into natural Canadian French: "Our office will be closed on Monday for Thanksgiving. Online orders will ship on Tuesday."
4. Rédige un courriel poli de quatre lignes pour reporter un rendez-vous à jeudi prochain.
5. Answer only in French: what is the capital of Australia, and why is it not Sydney?

Score each answer from 1 to 3 on three points: is it correct, is the language natural, and did it follow the instructions (length, language, tone)? A small scoring sheet makes it easy to compare sizes, or to compare Qwen with another local model you already use.

Try it: Run the five prompts with qwen2.5:3b and then qwen2.5 (7B). Score both, and note which mistakes disappear with the bigger model and which ones remain.

How good are they in English and French?

In English, Qwen 2.5 is strong for its size. Alibaba reports scores above 85 on MMLU (a broad general-knowledge test) and HumanEval (a programming test), and above 80 on MATH, for its best models. These are English-language benchmarks and refer to the larger versions, so expect less from the 3B and 7B models, though the team stresses that its small models have improved a great deal. The team also reports better instruction following, better handling of tables and more reliable JSON output, which matter for practical tasks.

French is on the official list of supported languages, but the announcement's headline numbers don't measure French writing, so the most useful evidence is your own test. Expect the larger sizes to do noticeably better than the small ones. Small models tend to slip in predictable ways in a second language: an anglicism, a stiff turn of phrase, a word from European French where a Quebec reader expects another (or the reverse), or a sudden switch back to English in a long reply. Ask for the variety you want ("in Quebec French") and check the output.

For anything published or sent to clients, treat a local model's French the way you would treat a quick first draft from a colleague: helpful for speed, but read by a fluent person before it goes out.