Qwen: the open model with the best multilingual support

If you want an open model that understands Italian and French without needing to translate, the simplest answer is Qwen. It is Alibaba's family of open source models and, of all the options you can run locally, it is the one that performs best in European languages. It is not the fastest, not always the smallest: it is the most capable where it matters for anyone writing in Italian, French, or other non-English languages.

In this guide I will explain how the Qwen family is built, why it handles Italian and French well, when it makes sense to choose it, and how it compares to Llama. By the end you will have a clear criterion for deciding which model to download.

The Qwen family: generations and sizes

Qwen is not a single model: it is a family in continuous evolution. The generations are recognized by the number in the version.

• Qwen2.5 — the consolidated generation. It is mature, well supported by every tool (Ollama, llama.cpp, LM Studio) and available in many sizes: 0.5B, 3B, 7B, 14B, 32B and 72B. If you want a stable and predictable model, start here.
• Qwen3 — the newest generation. It brings the MoE (Mixture-of-Experts) models and the thinking mode: the model can reason at length before answering. The MoE are the main novelty: they count many parameters in total, but activate only a small part of them per token. The result is the quality of a large model with the speed of a small one.

Sizes are read in parameters (B, billions). The higher number usually means more quality, but also more RAM and GPU required. Here is a practical map of what you can run where:

Model	Parameters	Reasonable minimum hardware
Qwen2.5 0.5B	0.5B	Any laptop, even CPU
Qwen2.5 3B	3B	8GB RAM, or a GPU with 6GB VRAM
Qwen3 4B	4B	16GB RAM / 8GB GPU
Qwen2.5 7B	7B	16GB RAM / 12GB GPU (quantized)
Qwen3 30B-A3B (MoE)	30B total, 3B active	16GB RAM / 12GB GPU
Qwen2.5 14B	14B	32GB RAM / 16GB GPU
Qwen2.5 32B	32B	64GB RAM / 24GB+ GPU
Qwen2.5 72B	72B	128GB RAM or 2x 48GB+ GPUs

The key point with MoE: the "parameters" column is misleading. Qwen3 30B-A3B has 30 billion parameters, but uses only 3 at a time, so it responds like a 3B model and weighs in memory like a 30B model. It is an excellent compromise: it belongs in the "minimum hardware" row lower than the number would suggest.

Why Italian and French come out well

Qwen's real strength is precisely multilingualism. Three concrete reasons.

1. Serious multilingual training. Qwen was trained with a high share of data in dozens of languages, not just English. Italian and French, being widely spoken languages, receive far more weight than models from other families that are "English and that's it".

2. A vocabulary that covers European languages better. Models with a rich vocabulary of multilingual tokens break Italian and French into more efficient units. Fewer tokens per word means more fluent answers and also faster generation, because the model produces fewer tokens for the same text.

3. Less of an "English accent". Many models, when answering in Italian, have a structure that is a bit too Anglophone: rigid sentences, odd word choices, calques from English. Qwen produces a more natural, idiomatic Italian, and equally a correct French. That is the difference between a machine translator and a native speaker.

A practical detail: if you use Open WebUI or LM Studio, set the language in the prompt ("Rispondi in italiano" / "Répondez en français"). Qwen obeys well, but giving it an explicit instruction removes any ambiguity, especially on the smaller models.

When to use it

Qwen is the right choice in these cases:

• You write or work in Italian or French and you want a model that does not force you to translate back and forth through English.
• You want a good quality/speed compromise without spending on top-tier hardware: this is where the MoE (30B-A3B) shine.
• You are building assistants, summarizing texts, drafting documents: the 7B to 32B models in Qwen2.5 cover these tasks well.
• You want to reason more on a complex problem: the thinking mode in Qwen3 is worth it, provided you have the hardware for the model.

When it is not the best choice: if you only need a 0.5B-3B model for fast runs on a modest PC and multilingual quality does not matter to you, the lighter and faster families are fine. And if your use case is purely in English and you want to squeeze every token of speed, it is worth comparing the alternatives too.

A quick comparison with Llama

Llama is the other great open family, from Meta. Here is the comparison that actually matters, without the marketing.

Aspect	Qwen	Llama
Multilingual (IT/FR)	Very strong	Decent, but more "English"
Vocabulary	Rich, efficient on European languages	More English-oriented
Reasoning mode	thinking (Qwen3)	present in recent versions
MoE	Yes (Qwen3)	No
Tool ecosystem	Universal (Ollama, LM Studio, llama.cpp)	Universal
License	Open, commercial use allowed	Open, commercial use allowed

In short: in English they are very close, in Italian and French Qwen usually wins. If your working language is English, Llama remains an excellent alternative. If you write in Italian or French, Qwen is the more sensible starting point. These are not two models to pick once and for all: it is worth trying both on your specific case, because the differences are felt mostly in your real texts, not in the benchmarks.

How to download and try it

With Ollama the test takes two minutes:

# Small, fast model, great to start
ollama run qwen2.5:3b

# Mid-size model, better quality
ollama run qwen2.5:14b

# MoE: high quality, high speed
ollama run qwen3:30b-a3b

Then ask the same thing in Italian and in English and compare the quality. An example test that gives you a real idea:

"Write a polite email to a customer to refund an order, in English, professional tone."

The model that gives you back a natural English text, with the right courtesies and no Anglophone construction, is the one for you. For Italian or French, run the same test with "Rispondi in italiano" or "Répondez en français".

If you want the maximum quality without hardware limits, there are also the larger Qwen models, but for most users a Q4-quantized 7B-14B, or the MoE 30B-A3B, is the sweet spot.

In summary

• Qwen is the open family with the best support for Italian and French, thanks to serious multilingual training and a vocabulary that renders European languages better.
• Two main generations: Qwen2.5 (mature and stable) and Qwen3 (with MoE and thinking mode).
• The MoE (30B-A3B) are the best quality/speed compromise: large-model quality, small-model speed.
• When to use it: you work in Italian/French, you want a good compromise on normal hardware, you build assistants and write documents.
• Versus Llama: close in English, Qwen usually wins in Italian and French.

The practical advice is always the same: download two sizes, test them on your real texts and choose with your own eyes. Local AI lets you do exactly that, without spending a single cent on a subscription.

Article generated with AI