Common problems & fixes
Aggregated common pitfalls (not an exhaustive catalogue):
- Garbled or over-verbose Instruct output. You are almost certainly not applying the chat
template - use
apply_chat_template, and don't prompt a base checkpoint as a chat model.
- Truncated or silently dropped context. The window is only 4k - long inputs overflow;
chunk or retrieve, and budget the 4k across system/input/output.
- Weak output in a specific language. Per-language quality is uneven across the 35
languages; evaluate your target language and consider a light fine-tune for lower-resource
ones.
- Non-reproducible results / silent updates. You floated on
main; pin an exact HF
revision (e.g. a dated -2512 snapshot) and verify checksums.
Versions, changelog & cadence
Releases follow a generational cadence - 1.7B → 9B → 22B - with dated snapshot revisions
(e.g. EuroLLM-9B-Instruct-2512) published on the utter-project org, each with base and
Instruct checkpoints. Individual versions are tracked as immutable Hugging Face revisions,
so the revision hash is your changelog anchor: pin it, and diff against a newer revision when
you choose to upgrade. The EuroLLM-9B technical report accompanies the release and documents
what changed.
- EuroLLM project site (
eurollm.io, ev-eurollm-site) for announcements and docs.
- Hugging Face model discussion tabs on the
utter-project org (ev-hf-9b-instruct) for
usage questions.
There is no paid support tier - this is community and consortium/maintainer support around an
EU-funded open project.
Tracked known issues
Drawn from technical-report caveats and third-party reviews rather than a formal issue tracker
(hence partial):
- Context window is only 4k - a real constraint for document workloads.
- Per-language quality is uneven; lower-resource EU languages trail the majors.
- Raw reasoning is in-class, not frontier - it is a 9B (or 1.7B) model.
- Base checkpoints are untuned research artifacts and must not be deployed as assistants.
The ownership factor this domain covers, drawn from the one entry record.
2
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
ModerateOpen-weights-plus-recipe, but only partial data openness: the technical report documents the tokenizer, architecture, data mixture and training procedure, and the released SFT dataset (EuroBlocks-SFT-2512) plus the EuroFilter artifact make the recipe inspectable - yet the ~4T-token pre-training corpus is described, not published as one dataset, so this is stronger than weights-only but weaker than a fully-open model (cf. OLMo) and not reproducible from a released corpus.
How this scores (AOI sub-dimensions)
Provenance3/5how well we can trace and verify what went into the modelDistributed from the verified utter-project org on Hugging Face in safetensors with per-file checksums (checklist ~5/8); the technical report makes the pipeline well-documented and the data sources are named.
Governance3/5how accountable and well-documented the publisher isAn EU-funded consortium with named academic and industrial partners, detailed technical reports and a clear release cadence (1.7B -> 9B -> 22B).
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.
Vendor announcementunverified2026-07-25
EuroLLM is presented as an open-source European multilingual LLM suite covering all 24 official EU languages plus additional strategic languages, developed by an EU-funded consortium.
Model cardread2026-07-25
EuroLLM-9B-Instruct model card, read verbatim: hosted on the verified utter-project org on Hugging Face in safetensors under Apache-2.0, trained on 4T tokens, instruction-tuned on EuroBlocks, a 4k context and a documented chat template (BOS id 1 / EOS id 4); noted as not preference-aligned (may hallucinate or produce harmful content); comparable to Gemma-2-9B on EU languages and Mistral-7B on English.
Model cardread2026-07-25
EuroLLM-1.7B-Instruct model card, read verbatim: 1.7B parameters, a 4,096-token context, coverage of 35 languages, Apache-2.0, and comparable to Gemma-7B on machine-translation benchmarks.
Technical_reportread2026-07-25
EuroLLM-9B: Technical Report (arXiv 2506.04079), read verbatim: EuroLLM is trained from scratch to cover all 24 official EU languages plus 11 more, and documents the tokenizer, architecture, ~4T-token data mixture, EuroFilter pipeline, EuroBlocks-Synthetic post-training set and Megatron-LM training; the report frames its "Open release" as the "Public availability of models, filters, and datasets."
Documentationread2026-07-25
EuroLLM-22B-Instruct-2512 model card, read verbatim: openness is PARTIAL, not fully-open - the SFT dataset EuroBlocks-SFT-2512 IS released (utter-project/EuroBlocks-SFT-2512), but the ~4T-token pretraining corpus (web, parallel, Wikipedia, Arxiv, books, math, code, Apollo) is only described, not published as one downloadable dataset.
Vendor announcementunverified2026-07-25
The EuroLLM team's release blog describes EuroLLM-9B, its EU consortium and EU funding (EuroHPC / Horizon Europe), multilingual coverage, Apache-2.0 licensing and the release of the models, filter and synthetic post-training data.
Licenceread2026-07-25
EuroLLM base and Instruct weights - 9B, 1.7B and 22B-Instruct-2512 - are distributed under OSI-approved Apache-2.0 with unconditional commercial use and no field-of-use limits (the licence is not a preference-alignment/safety grant, and the cards warn the models may hallucinate).
Third-party analysisunverified2026-07-25
The European Commission's Open Source Observatory profiles EuroLLM as EU-funded pioneering European open-source AI released under an open licence.
Third-party analysisunverified2026-07-25
An independent review of the EuroLLM-9B technical report reports it on par with Gemma-2-9B on multilingual benchmarks, ahead on WMT24++ translation, and matching Mistral-7B on English.
Third-party analysisunverified2026-07-25
Independent coverage ranks EuroLLM among the leading European-developed open models for multilingual EU-language performance.
Third-party analysisunverified2026-07-25
EuroLLM safetensors checkpoints load on mainstream serving stacks (vLLM, llama.cpp, Ollama, TGI, transformers), with community GGUF/MLX/GPTQ quants and a hosted NVIDIA NIM endpoint available.