EuroLLM 9B Instruct
A 9B-parameter model from the EU-funded EuroLLM project, instruction-tuned for chat and translation in all 24 official EU languages plus 11 others, among them Turkish and Arabic.
Overview
EuroLLM-9B-Instruct is the instruction-tuned version of EuroLLM-9B, a 9-billion-parameter dense transformer from the EU-funded EuroLLM project. Nine organizations built it, among them Unbabel, Instituto Superior Técnico, the University of Edinburgh and Sorbonne Université.
The base model was pre-trained on 4 trillion tokens. Alongside web text and high-quality datasets, the mix included parallel data pairing English with each of the other languages in both directions. The instruct model was then tuned on EuroBlocks, a dataset aimed at general instruction following and machine translation.
It covers the 24 official EU languages, small ones like Irish and Maltese included, and 11 more: Arabic, Catalan, Chinese, Galician, Hindi, Japanese, Korean, Norwegian, Russian, Turkish and Ukrainian.
The EuroLLM team's multilingual comparison mixes five benchmarks run on the pre-trained model with two translation benchmarks (WMT24 and FLORES) run on the instruct version. Across them it reports a Borda count of 1.0 for EuroLLM-9B against other European-developed open models, meaning it ranked first on average, and calls its results comparable to Gemma-2-9B. On English benchmarks it reports the same Borda count as Mistral-7B.
What you can build
- Translating product copy and help-center articles between English and any of the other 34 languages.
- Support assistants that reply to each customer in their own language and remember earlier turns.
- Question answering over short retrieved passages in European languages.
- First drafts and summaries in smaller EU languages such as Slovenian or Irish.
Inputs
Only the prompt is required. An optional system prompt is prepended to it to set the assistant's role and tone; the default is a generic helpful-assistant instruction.
For multi-turn chat, send the same user_id and session_id with each request and the earlier turns of that session are used as history. One user_id can hold several sessions. Leave both blank for one-off requests.
Temperature defaults to 0.7 (range 0 to 5), top_p to 0.95, and top_k accepts 0 to 100. Repetition and length penalties both start at 1. max_tokens and min_tokens bound the reply length, and stop_sequences takes a semicolon-separated list. The seed defaults to 123456. Quantization is on by default to reduce memory use. do_sample is on too; switch it off for deterministic decoding.
Outputs
A plain text reply. The sample on this page asks about events near the Tower of London. The one-paragraph answer names Anne Boleyn's execution in 1536 and two prisoners held there, Sir Walter Raleigh and Rudolf Hess, and suggests loyalty and ambition as story themes.
Recommended settings
- For translation, turn do_sample off and name the source and target language in the prompt.
- For open-ended chat, start from the defaults.
- Put a glossary or terminology rules in the system prompt so every request carries them.
- If long replies start looping, raise the repetition penalty slightly above 1.0.
Limitations
- The card's hyperparameter table gives a sequence length of 4,096 tokens. Chat history and pasted context count toward it, and so does the reply.
- Most of the card's benchmark scores were measured on the pre-trained EuroLLM-9B. Only the WMT24, FLORES and IFEval columns were run on this instruct version.
Safety & compliance
EuroLLM-9B-Instruct is released under the Apache License 2.0. The model card states that it has not been aligned to human preferences, so its output can include hallucinations and other false statements as well as harmful content.
API quick start
Run EuroLLM-9B-Instruct with a single API call.
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/utter-project/eurollm-9b-instruct" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "What are some interesting historical even…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



