Qwen vs Llama vs Gemma: which open model for Spanish and European languages

What each model card says about languages

The first thing to know is that “supported” does not mean the same thing on every card. Meta lists a closed set of languages for each Llama release. Alibaba and Google give a count. The BSC publishes the share of each language in its training data. Here is the wording from each official card:

Checked on the official model cards and policies on 9 November 2026. Sources: huggingface.co/Qwen/Qwen3.5-9B, huggingface.co/Qwen/Qwen3-14B, huggingface.co/Qwen/Qwen3.8-27B, huggingface.co/meta-llama/Llama-3.3-70B-Instruct, huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct, huggingface.co/google/gemma-4-26B-A4B-it, huggingface.co/BSC-LT/ALIA-40b-instruct-2606

Meta’s lists matter more than they look. Both Llama cards put use “in languages beyond those explicitly referenced as supported” under out-of-scope use. The same cards say the models were trained on a broader set of languages and that developers may fine-tune them for others, as long as they comply with the licence and the acceptable use policy. In practice, outside those languages you are on your own.

Qwen and Gemma go the other way: a large count, no per-language list and no per-language quality figures on the card. Broad coverage is a good sign; it is not a guarantee for any single language.

Spanish: all three support it, so decide on something else

Spanish is on every list above, so for Spanish alone the language question does not separate them. The multilingual benchmarks the cards publish point the same way, but they are different tests on models of different sizes, so read them as “all good”, not as a ranking:

Checked on the official model cards and policies on 9 November 2026. Sources: huggingface.co/Qwen/Qwen3.5-9B, huggingface.co/google/gemma-4-26B-A4B-it, huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct

So pick on the things that do differ: the licence (next sections), the size you can afford to run, and what else the model must do. If your Spanish-language product also needs tool calling for an agent, Qwen3.8 27B and Qwen3.5 9B have it. If it reads screenshots or photos, Gemma 4 does. If it is a high-volume chat in Spanish, a small model such as Qwen3.5 9B is usually enough, and you should test it on your own conversations before paying for a bigger one.

French, German, Italian, Portuguese and the rest of Europe

Llama 4’s official list includes six European languages: English, French, German, Italian, Portuguese and Spanish. Llama 3.3 has the same six. Dutch, Polish, Swedish, Czech, Greek, Romanian and the other EU languages are not on either list, which under Meta’s own card makes them out-of-scope use.

Qwen and Gemma claim much wider coverage. The closest thing to a Europe-wide figure on these cards is Qwen3.5 9B’s MMLU-ProX score, averaged over 29 languages, and its WMT24++ translation score, averaged over 55. Gemma 4’s card says it works out of the box in more than 35 languages and saw more than 140 in pre-training.

ALIA’s base model was pre-trained on 35 European languages, from Bulgarian and Czech to Irish, Maltese and Ukrainian. Its instruction tuning, though, focused on Spanish, Catalan, Basque, Galician and English, so outside those five treat it like any broad multilingual model: test before you rely on it.

Catalan, Basque and Galician: where ALIA comes in

None of the Llama cards lists Catalan, Basque or Galician, and the Qwen and Gemma cards do not break their counts down by language. ALIA 40B is the one model in this comparison that was built around them. The BSC doubled the weight of Spanish and the co-official languages in the training data; its base-model card gives the final shares:

Checked on the official model cards and policies on 9 November 2026. Sources: huggingface.co/BSC-LT/ALIA-40b, huggingface.co/BSC-LT/ALIA-40b-instruct-2606

The BSC also publishes base-model evaluations in each of the four, from Catalan commonsense reasoning to Basque exam questions. Two limits from the instruct card: ALIA does not call functions or external APIs (the BSC publishes a separate function-calling variant for that), and it is text only. For an agent that must act, pair it with a Qwen model for the tool calls and let ALIA write what the user reads.

The licence question for European companies

Qwen, Gemma 4 and ALIA are Apache 2.0: commercial use, modification and fine-tuning allowed, with the notices kept. That is the simple case.

Llama 4 is not. Its acceptable use policy says that, with respect to any multimodal models included in Llama 4, the rights granted under section 1(a) of the Llama 4 Community License “are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union.” The policy adds that this does not apply to end users of a product or service that incorporates those models. The Llama 4 card describes the collection as natively multimodal, so a company headquartered in the EU should read that clause with its lawyers before building on Llama 4.

The Llama 4 licence also has a size threshold: if the products or services of the licensee and its affiliates had more than 700 million monthly active users in the calendar month before the release date, it must request a licence from Meta, which Meta may grant or refuse.

One more data point: Meta’s open-weight release of August 2026, Muse Glimmer 30B, is not branded Llama and ships under Apache 2.0. Its card says it was trained on data from more than 100 languages and has not been evaluated on all of them, with no per-language list.

Checked on the official model cards and policies on 9 November 2026. Sources: dev.meta.ai/llama/llama4/use-policy/, huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct, huggingface.co/meta-models/Muse-Glimmer-30B

Which one should you pick?

Putting the cards and the licences together, this is the short version for a team building for Spain or the rest of Europe:

How this works on GPU Flow

We do not serve Llama. From the families in this article, Token Factory serves Qwen, Gemma and ALIA through an OpenAI-compatible API at https://api.gpuflow.ai/v1, billed per million tokens in euros from a prepaid balance. These are today’s prices, from the same source as our pricing page, with VAT included:

Model pages: Qwen3.8 27B, Qwen3 14B, Qwen3.5 9B, ALIA 40B Instruct, Gemma 4 26B A4B, Qwen3-VL 32B OCR

ALIA 40B runs unquantised (BF16) on a full NVIDIA B200; the smaller Qwen and Gemma models run on MIG slices of one, which is why their served context can be shorter than the native one. Switching between them is a change of the model parameter, so you can send the same Spanish or Catalan prompt to three models and compare the answers.

New accounts get 5 € of free Token Factory credit for exactly that kind of test. A card is required to sign up, and when the balance runs out the API returns an error instead of charging the card. Nothing you send is used to train any model.

Related

• Best open-source LLMs in 2026 for business use: Qwen3.8, GLM 5.3, Gemma 4 and Nemotron comparedQwen3.8, GLM 5.3, Gemma 4 and Nemotron compared on licence, languages, context and the benchmarks on their official model cards, with live API prices.
• Migrate from the OpenAI API to an open-source alternative in 15 minutes (code included)Move from the OpenAI SDK to GPU Flow in five steps: change the base URL, pick the equivalent model and keep streaming, tools and reasoning.