BLOG · COMPARISON · · UPDATED

Open LLMs for German, Polish and Romanian: European models, licences and GPU memory compared

Eurokommerz, Vienna, since 2006: Private AI/ML · IT Managed Services · Enterprise Training · AI Hardware & Software

IN BRIEF
  • EuroLLM (1.7B, 9B and 22B, Apache 2.0) and Teuken-7B name German, Polish and Romanian among all 24 official EU languages; Teuken’s commercial v0.4 is Apache 2.0, its newer v0.6 CC BY-NC 4.0 for non-commercial use, and Aleph Alpha’s Kolibri-1 (3 October 2026, Apache 2.0) covers German and English
  • For Polish, Bielik v3 (1.5B to 11B, Apache 2.0) and PLLuM (4B and 12B under Apache 2.0, 8B and 70B under Meta’s Llama 3.1 licence) are trained with an emphasis on the language; the Romanian instruct models of OpenLLM-Ro are CC BY-NC 4.0
  • Mistral Small 3.1, 3.2 and 4 name all three languages and are Apache 2.0, while Mistral Medium 3.5’s modified MIT licence grants no rights to companies above a monthly revenue threshold
  • Qwen3’s launch post lists all three languages among “119 languages and dialects”, Gemma 4’s card gives 35+ languages out of the box and 140+ in pre-training without naming them, and Llama 3.1, 3.3 and 4 list German but not Polish or Romanian; OpenEuroLLM had published only research checkpoints as of 23 October 2026
  • In BF16, EuroLLM-22B takes 42.2 GiB, which leaves 40.9 GiB of KV cache on one 96 GB RTX PRO 6000 by our users-per-card rule, enough for about 24 conversations of 8,192 tokens

Eurokommerz × Vixen.UNO: Private AI/ML  Talk to an expert →

Open LLMs for German, Polish and Romanian: what the model cards list

The open models whose cards name German, Polish and Romanian include EuroLLM (1.7B, 9B and 22B), Teuken-7B, the 7B and 11B Bielik v3 models and Mistral Small 3.1, 3.2 and 4, and Qwen3’s launch post lists all three among 119 languages and dialects (the Qwen3-14B card says “100+ languages and dialects”). Bielik and PLLuM are trained with an emphasis on Polish, and the OpenLLM-Ro project publishes Romanian fine-tunes for research use. EuroLLM, Bielik v3 and these Mistral Small versions are Apache 2.0, while Teuken’s current version and the OpenLLM-Ro instruct models exclude commercial use and two PLLuM sizes come under Meta’s Llama 3.1 licence. “Open” here means downloadable weights under the licence on the card, which is not always an open source licence: the Open Source Definition does not allow a licence to restrict use “in a business”. We supply the GPU servers for models such as these, and our Private AI/ML service deploys them on-premise. Write to us about your languages and users and our team will help.

MODEL, ORIGINSIZESDE, PL, ROLICENCE ON THE CARD
EuroLLM (EU-funded)1.7B, 9B, 22B; EuroMoE 2.6Ball three, among 35 languagesApache 2.0
Teuken-7B (OpenGPT-X)7Ball three, among the 24 EU languagesv0.4 commercial: Apache 2.0; v0.6: CC BY-NC 4.0
Kolibri-1 (Aleph Alpha)78B MoE, 3.46B activeGerman named; Polish, Romanian notApache 2.0
Salamandra, ALIA (BSC)2B, 7B; ALIA 40Ball three in pre-training; instruct tuning mainly Iberian languages and EnglishApache 2.0
Bielik v3 (SpeakLeash)1.5B, 4.5B, 7B, 11B7B, 11B: all three, among 32 European languages; 1.5B, 4.5B: PolishApache 2.0; contact details to download
PLLuM (HIVE AI)4B, 8B, 12B, 70BPolish, with English data4B, 12B: Apache 2.0; 8B, 70B: Llama 3.1 licence
Mistral Small 3.1, 3.2, 424B; 119Ball three namedApache 2.0
Mistral Medium 3.5128Ball three namedmodified MIT with a revenue threshold
Ministral 3, Mistral Large 33B to 14B; 675BGerman named; Polish, Romanian notApache 2.0
Apertus (Swiss AI)0.5B to 70B“over 1000 languages”, none namedApache 2.0 and an acceptable use policy
OpenLLM-Ro instruct models2B to 9BRomanianCC BY-NC 4.0

Hugging Face model cards and language metadata, read on 6 October 2026 (Kolibri-1 on 23 October 2026). DE, PL, RO: whether the card or its metadata names German, Polish and Romanian.

EuroLLM, Teuken and OpenEuroLLM: models for all 24 EU languages

EuroLLM is developed by ten universities, research centres and companies with EU funding, and its cards list the 24 official EU languages and 11 more. EuroLLM-22B has 22.6 billion parameters and a context of 32,768 tokens. Its card says it matches or outperforms Gemma-3-27B, Qwen-3-32B and Apertus-70B in translation. The instruct cards of the 9B and 22B models state that the model “has not been aligned to human preferences” and may produce hallucinations or false statements.

Teuken-7B was developed in the research project OpenGPT-X by Fraunhofer, Forschungszentrum Jülich, TU Dresden and DFKI. It covers the 24 official EU languages with a context of 4,096 tokens, which leaves little room for retrieved passages in a RAG prompt. Teuken-7B-instruct-commercial-v0.4 is Apache 2.0 and “intended for commercial and research use”, while the newer v0.6 is CC BY-NC 4.0 and “designed for private, non-commercial, research, and educational use”. Both cards say the model is “not intended for use in math and coding tasks”. The German company Aleph Alpha released Kolibri-1 on 3 October 2026 for German and English under Apache 2.0. It is a mixture-of-experts model with 78.1B parameters, 3.46B of them active, about 78 GB in FP8, with 1,048,576 tokens of context advertised and up to 262,144 recommended. Its card lists one H200 or two 80 GB A100 among the minimum setups. By our users-per-card rule below, the FP8 checkpoint, 78.8 GB or 73.4 GiB in 32 files on Hugging Face, leaves about 9.6 GiB of cache on one RTX PRO 6000 and about 93 GiB on two; the card names no RTX PRO GPU, so check that your serving engine runs its block-scaled FP8 format on these cards. Our guide to Kolibri-1 hardware requirements sizes it on the RTX PRO 6000 Server Edition and the H200 NVL.

OpenEuroLLM started on 1 February 2025 with 20 organisations and funding from the Digital Europe Programme, to build open models for the “EU official languages and beyond”. As of 23 October 2026 its Hugging Face organisation still holds only research checkpoints, the newest from 21 October, and on 6 October the card of its instruction-tuned 9B model called it an “Experimental SFT-stage research checkpoint; not a final production assistant”. In March 2026 the project wrote that it “is looking to release an 8B model by next summer”; no such release had appeared by 23 October 2026.

Salamandra, ALIA and Apertus: what pre-training and tuning cover

Salamandra, from BSC, the supercomputing centre in Barcelona, comes in 2B and 7B sizes pre-trained on 35 European languages and code, German, Polish and Romanian among them. ALIA-40b was pre-trained on 9.83 trillion tokens of 35 European languages, and all of these models are Apache 2.0. ALIA-40b-instruct-2606 states that “the post-training process concentrated primarily on Spanish, Catalan, Basque, Galician, and English”, and its metadata, like that of salamandra-7b-instruct-2606, lists only those five languages. Test these instruct models in German, Polish or Romanian before you rely on them.

The Swiss AI Initiative describes Apertus as a “Fully Open Model” with open weights and open training data, and its 2509 card says it “supports over 1000 languages”. Version 1.5, in 8B and 70B sizes, is Apache 2.0, but downloading it means accepting an acceptable use policy and sharing contact details; version 1.1 adds 0.5B, 1.5B and 4B models. The card advises fetching new weights every six months, as the project handles data-protection deletion requests through new releases.

Polish LLMs: Bielik and PLLuM

SpeakLeash and ACK Cyfronet AGH develop Bielik. In its v3 generation the 1.5B and 4.5B models are tagged for Polish only, while the 7B and 11B models are “Multilingual (32 European languages, optimized for Polish)”, with German and Romanian in their metadata. All are Apache 2.0, and the download page asks you to agree to share contact information. The 11B instruct card notes that the model “does not have any moderation mechanisms”. Bielik-PL-11B-v3.0-Instruct uses the APT4 tokenizer, “optimized specifically for the Polish language”, and Bielik-Minitron-7B is an instruct model built on a pruned version of the 11B base model. SpeakLeash also publishes FP8, AWQ and GGUF versions of the 11B instruct model.

PLLuM is, in its card’s words, “a family of large language models (LLMs) specialized in Polish with additional English data incorporated for broader generalization”, developed by the PLLuM consortium in 2024 and since 2025 under HIVE AI, an initiative for the Polish public sector financed by Poland’s Minister of Digital Affairs. The December 2025 generation, marked 2512, comes in base, instruct and chat variants. PLLuM-4B is built on Google’s gemma-3-4b-pt and PLLuM-12B on Mistral-Nemo-Base-2407, both Apache 2.0 on their cards. Llama-PLLuM-8B and Llama-PLLuM-70B are built on Llama 3.1 and come under Meta’s Llama 3.1 licence and its acceptable use policy; the licence requires anyone who distributes them, or a product or service that contains them, to display “Built with Llama”. Google’s Gemma Terms of Use count “works based on Gemma” as model derivatives and pass their use restrictions on, so how they combine with the Apache 2.0 on the PLLuM-4B card is a question for your legal department.

Romanian LLMs and Romanian in multilingual models

Of the models above, EuroLLM, Teuken, the Salamandra base models, Bielik’s 7B and 11B models, Mistral Small 3.1, 3.2 and 4, Mistral Medium 3.5 and Qwen3 name Romanian, while Llama 3.1, 3.3 and 4 do not. The OpenLLM-Ro project fine-tunes Llama, Mistral, Gemma and Qwen models for Romanian and publishes its instruct models under CC BY-NC 4.0, which excludes commercial use. The RoLlama3.1-8b-Instruct card, version of 23 April 2025, describes the model as intended for research use in Romanian. It reports 6.43 on RoMT-Bench, against 5.69 for the Llama 3.1 8B Instruct it starts from.

Mistral open-weight models: one licence per model

Mistral AI publishes most of its open-weight models under Apache 2.0, among them Mistral NeMo, Mistral Small 3.1 and 3.2 (24B), Mistral Small 4 (119B), the Ministral 3 models (3B, 8B and 14B) and Mistral Large 3 (675B). The Mistral Small 3.1 card names 25 languages, German, Polish and Romanian among them, and the 3.2 card calls that model “a minor update” of 3.1. The newer cards name German among “dozens of languages”; Polish and Romanian appear only in the metadata of Mistral Small 4 and Mistral Medium 3.5, not in that of Ministral 3 or Mistral Large 3. Mistral announced Mistral Large 4 (1T parameters, 52B active) on 6 October 2026 and published the weights on Hugging Face on 29 October 2026, under the Mistral Ecosystem Open-Weight License, a modified MIT licence that requires a separate commercial licence for Model-as-a-Service offerings.

Mistral Medium 3.5, a dense 128B model, comes under a “Modified MIT License” that grants no rights at all if the “global consolidated monthly revenue” of your company exceeds a threshold set in the licence, and the condition extends to derivatives. Older models such as Mistral Large 2411 and Ministral 8B 2410 use the “Mistral Research License”, which allows non-commercial use; Mistral offers a commercial licence on request.

Qwen3, Gemma 4, Llama and gpt-oss: languages on the cards

MODELWHAT THE CARD SAYSDE, PL, ROLICENCE
Qwen3“100+ languages and dialects”; 119 in the launch postall three in the launch postApache 2.0 (Qwen3-32B)
Qwen3.8no language list on the 27B card (August 2026)none namedApache 2.0 (Qwen3.8-27B)
Gemma 435+ out of the box, 140+ in pre-trainingnone namedApache 2.0
Llama 3.1, 3.3eight supported languagesGerman onlyMeta’s community licence and use policy
Llama 412 supported, 200 in pre-trainingGerman onlyLlama 4 licence; multimodal rights withheld in the EU
gpt-ossno language listGerman tested in MMMLUApache 2.0 and a usage policy

Model cards; Qwen3 launch post (29 April 2025); Meta’s Llama 4 acceptable use policy; OpenAI’s gpt-oss model card (5 August 2025); all read on 6 October 2026.

OpenAI’s gpt-oss model card of 5 August 2025 reports MMMLU, a professionally translated MMLU in 14 languages without Polish or Romanian, on which gpt-oss-120b scores 83.0 in German at high reasoning effort. Our guide to a private ChatGPT alternative compares their licence terms, including Llama 4’s clause on companies with their principal place of business in the EU. Meta’s newer open line, Muse, includes Muse Glimmer 30B under Apache 2.0 (August 2026); check the languages on its card before you shortlist it.

GPU memory per model size

Our users-per-card rule takes 0.9 of the memory the driver reports and subtracts about 3 GiB of overhead and the weights, two bytes per parameter in BF16; the rest holds the KV cache, as our guide to how much VRAM an LLM needs explains.

MODELBF16 WEIGHTSRTX PRO 5000, 48 GBRTX PRO 6000, 96 GBH200 NVL, 141 GB
EuroLLM-9B-Instruct-251217.0 GiB23.2 GiB66.0 GiB106.3 GiB
Bielik-11B-v3.0-Instruct20.8 GiB19.4 GiB62.2 GiB102.6 GiB
PLLuM-12B-instruct-251222.8 GiB17.4 GiB60.2 GiB100.5 GiB
EuroLLM-22B-Instruct-251242.2 GiBdoes not fit40.9 GiB81.2 GiB
Mistral-Small-3.2-24B44.7 GiBdoes not fit38.3 GiB78.6 GiB
ALIA-40b-instruct-260675.3 GiBdoes not fit7.7 GiB48.0 GiB
Llama-PLLuM-70B-instruct131.4 GiBdoes not fitdoes not fitdoes not fit

Cache room per card by this rule, with 95.6 GiB (RTX PRO 6000) and 140.4 GiB (H200 NVL) as the driver reports them and 48 GiB for the RTX PRO 5000, for which we found no driver figure; parameter counts from Hugging Face, 6 October 2026. “Does not fit” means no room left for the cache. NVIDIA’s vPC sizing guide, updated on 19 August 2026, states that Blackwell-generation GPUs use ECC that is built into GDDR7 memory and always enabled without reducing usable VRAM. We size with the memory the driver reports (nvidia-smi, FB Memory Total) where we have it.

EuroLLM-22B stores 216 KiB of cache per token in BF16, from the 54 layers and 8 KV heads of 128 dimensions in its config.json; the card’s table gives 56 layers, which does not match the parameter count. The 40.9 GiB left on one RTX PRO 6000 therefore hold about 24 conversations of 8,192 tokens, or 6 at the full 32,768, and an FP8 cache doubles both. Llama-PLLuM-70B needs two RTX PRO 6000 cards in BF16, which leave about 17 GiB of cache per card, or a quantised version; our guide to FP8, NVFP4, INT4 and GGUF sets out what each format saves and what it costs in accuracy.

We build AI servers to order with the RTX PRO 5000, 6000 or H200 NVL; four RTX PRO 6000 Server Edition cards hold four shortlisted models of up to 24B side by side for a test. Send us your shortlist and your number of users.

Evaluating a model in your language with your documents

Public benchmarks contain none of your contracts or support tickets, so a test of your own decides between two or three candidates.

  1. Shortlist models whose licence covers your intended use and whose weights fit your GPUs at the precision you plan to run.
  2. Build the evaluation set described in our article on RAG on company data in every language your users write, with questions in one language about documents in another.
  3. Serve each candidate with the same engine, prompt, retrieved passages and context limit, for example through vLLM’s OpenAI-compatible API; our comparison of serving engines covers the options.
  4. Have native speakers grade the answers blind for correctness against the source, the language of the answer, grammar, terminology and the citation.
  5. Count the tokens each tokenizer produces for a sample of your documents, since the counts differ between models, and measure time to first token on your hardware.
  6. Run the set again on each new version; EuroLLM, Bielik and PLLuM have each published several generations.

The technical assessment in our Private AI/ML service covers model and GPU selection and a pilot plan with metrics, at a price fixed before work begins. Tell us which languages and document types your assistant has to handle.

What we do

Our Private AI/ML service deploys open and commercial models on-premise with vLLM, Ollama or NVIDIA AI Enterprise, so data does not leave your network and access is controlled, or on dedicated hardware in a Tier III data centre in Lithuania, where data stays in the EU. The first call is free of charge and leaves you with two or three possible solution scenarios. We start with a pilot on one process with clear metrics and scale only what has proved its value. Eurokommerz holds the contract and supplies the GPU servers, with engineering by our partner Vixen.UNO and support under an agreed SLA.

FAQ

What is PLLuM?
PLLuM is a family of open LLMs specialised in Polish, with additional English data, developed in 2024 by the PLLuM consortium and since 2025 under HIVE AI, an initiative financed by Poland’s Minister of Digital Affairs. The December 2025 generation comes in 4B, 8B, 12B and 70B sizes, in base, instruct and chat variants (the 70B as instruct and chat only), built on Gemma 3 4B, Llama 3.1 8B, Mistral NeMo and Llama 3.1 70B. The 4B and 12B models are published under Apache 2.0, the Llama-based 8B and 70B models under Meta’s Llama 3.1 licence.
Which open LLMs support German?
EuroLLM, Teuken-7B, Mistral Small 3.1 and 3.2 and the 7B and 11B Bielik v3 models name German on their cards, as do Llama 3.1, 3.3 and 4 and the newer Mistral models, and Qwen3’s launch post lists it among 119 languages and dialects. Aleph Alpha’s Kolibri-1, released on 3 October 2026 under Apache 2.0, covers German and English. Salamandra and ALIA were pre-trained on German, but the post-training of their current instruct versions concentrated on Spanish, Catalan, Basque, Galician and English. Which of them answers well enough is a matter for a test with your own questions, and published scores such as the 83.0 that OpenAI reports for gpt-oss-120b on German MMMLU are the developers’ own.
Which open LLMs are built for Polish?
Bielik, developed by SpeakLeash and ACK Cyfronet AGH, and PLLuM, developed under the Polish HIVE AI initiative, are both trained with an emphasis on Polish. Bielik v3 comes in 1.5B, 4.5B, 7B and 11B sizes under Apache 2.0, and PLLuM in 4B and 12B under Apache 2.0 and in 8B and 70B under Meta’s Llama 3.1 licence. EuroLLM, Teuken-7B, Mistral Small 3.1, 3.2 and 4 and Qwen3 also list Polish among their languages.
Is there an open LLM for Romanian?
The OpenLLM-Ro project publishes Romanian fine-tunes of Llama, Mistral and Gemma models, such as RoLlama3.1-8b-Instruct, under CC BY-NC 4.0, which excludes commercial use. For commercial use, EuroLLM, Teuken-7B commercial v0.4, the Bielik v3 7B and 11B models and Mistral Small 3.1, 3.2 and 4 name Romanian on their cards under Apache 2.0, and Qwen3 lists it in its launch post. Test the candidates in Romanian with your own questions before you choose.
What is EuroLLM?
EuroLLM is a family of open LLMs trained from scratch on all 24 official EU languages and 11 more, developed by universities, research centres and companies with EU funding. It comes in 1.7B, 9B and 22B sizes and as a 2.6B mixture-of-experts model, all under Apache 2.0, and the 9B and 22B models of the 2512 release accept 32,768 tokens of context. The 9B and 22B cards state that the models have not been aligned to human preferences and may produce hallucinations or false statements.
Which European open LLMs can be used commercially?
EuroLLM, Teuken-7B commercial v0.4, Kolibri-1, Salamandra, ALIA, Bielik v3, PLLuM-4B, PLLuM-12B and most open-weight Mistral models are published under Apache 2.0, and Apertus under Apache 2.0 with an acceptable use policy. Teuken-7B v0.6 and the OpenLLM-Ro instruct models are CC BY-NC 4.0, the Llama-based PLLuM models come under Meta’s Llama 3.1 licence, and Mistral Medium 3.5’s modified MIT licence grants no rights to companies above a monthly revenue threshold. Terms are set per model and version, so have your legal department read the terms of the exact checkpoint you deploy.

Our engineering partner helps you select open models and deploys them on-premise; we build the GPU servers to order. Describe your task in a few lines and our team will help: we reply within one business day. The first call is free of charge.

Talk to an expert
Talk to an expert

We reply within one business day

By sending this form you agree that we process your details to answer your enquiry; see our privacy policy.

request@eurokommerz.at
Jordangasse 7, 1010 Vienna