from mlx_lm import load, generate

model, tokenizer = load(
    "utter-project/EuroLLM-22B-Instruct-2512",
    adapter_path="Ailiance-fr/eurollm-multilingual-eu-lora",
)

print(generate(model, tokenizer, prompt="..."))

Training

Hyperparameter	Value
Base model	`utter-project/EuroLLM-22B-Instruct-2512`
Method	LoRA via `mlx-lm`
Rank	16
Scale	2.0
Alpha	32
Max seq length	2048
Iterations	500
Optimizer	Adam, LR 1e-5
Hardware	Apple M3 Ultra 512 GB

Training data lineage

Derived from the internal eu-kiki / mascarade curation. All upstream samples are synthetic, permissively-licensed, or generated from Apache-2.0 base resources. See the Ailiance-fr catalog for related cards.

Benchmark roadmap

This LoRA has not yet been evaluated through electron-bench (the current pipeline supports gemma-4-E4B base only). Training was completed with the standard mlx-lm LoRA trainer (rank 16, alpha 32, scale 2.0, AdamW LR 1e-5, 500 iters) — full hyperparameters are in the Training table above.

Planned evaluations:

Perplexity on the validation split of the training data
Functional benchmark on eurollm-specific tasks
Comparison vs base utter-project/EuroLLM-9B-Instruct

Track progress: ailiance-bench issues.

For reference benchmarks on the gemma-4-E4B base, see the base-vs-LoRA matrix.

License chain

Component	License
Base model (`utter-project/EuroLLM-22B-Instruct-2512`)	apache-2.0
Training data (internal Ailiance curation (synthetic + permissive sources))	apache-2.0
LoRA adapter (this repo)	apache-2.0

All upstream components are Apache 2.0 / MIT — LoRA inherits permissive terms.

EU AI Act compliance

Article 53(1)(c): training data licenses preserved (per-dataset cards declare upstream licenses).
Article 53(1)(d): training data summary — see upstream dataset cards on Ailiance-fr.
GPAI Code of Practice (July 2025): base utter-project/EuroLLM-22B-Instruct-2512 released under apache-2.0.
No web scraping by Ailiance, no licensed data, no PII.
Upstream Stack Exchange content (where applicable) is CC-BY-SA-4.0 and propagates to this adapter.

License

LoRA weights: apache-2.0 — see License chain table above for derivation rationale.

Citation

bibtex

@misc{ailiance_eurollm_multilingual_eu_2026,
  author    = {Ailiance},
  title     = {Ailiance — EuroLLM-22B-Instruct multilingual-eu LoRA},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/Ailiance-fr/eurollm-multilingual-eu-lora}
}

See the full Ailiance-fr LoRA collection.

Upstream base model — official evaluations

This LoRA fine-tunes utter-project/EuroLLM-22B-Instruct-2512, the best EU-made fully-open LLM at its scale (per the upstream model card, 2026 release). Architecture: 22.6B params (21.07B non-embedding), 56 layers, GQA (48 heads / 8 KV heads), 32k context, RoPE Θ=1M.

Official benchmark coverage (per EuroLLM-22B Technical Report):

Track	Benchmarks
Multilingual	HellaSwag · MMLU · MMLU-Pro · ARC-Challenge · MGSM · FLORES · WMT24++
English	IFEval · HellaSwag · MMLU · MMLU-Pro · BBH · ARC-Challenge · GPQA · GSM8K · MATH-500 · HumanEval
Translation	FLORES, WMT24++ across all 24 official EU languages

Per the official card: "The model excels at translation tasks being capable of translating across all official EU languages, matching or outperforming strong models like Gemma-3-27B, Qwen-3-32B and Apertus-70B. Furthermore, when it comes to general benchmarks, it is the best EU-made fully open model."

Full numbered tables (rendered as figures in the upstream card) and Borda Count rankings are in the Technical Report.

Source: official EuroLLM-22B-Instruct-2512 model card.

Reading these alongside this LoRA: EuroLLM-22B was designed for EU-language multilingual coverage and matches Apertus-70B / Qwen-3-32B on general benchmarks at less than 1/3 the size. This LoRA inherits the multilingual EU coverage and adds the domain specialization.

Ailiance-fr/eurollm-multilingual-eu-lora

Ailiance — EuroLLM-22B-Instruct multilingual-eu LoRA

Quick start (MLX)