Iris-300M
Iris-300M is a 290.6M-parameter encoder-decoder Transformer for machine translation between 11 European languages, trained from scratch on a single RTX 3070. It translates directly between any pair of supported languages (no English pivot).
Part of the Iris model family by BranchingNLP (Michelangelo Di Nicola).
Supported languages
enEnglishitItalianfrFrenchesSpanishptPortuguesedeGermannlDutchroRomanianplPolishcsCzechsvSwedish
Architecture
| Parameters | 290,641,920 |
| Encoder / decoder layers | 10 / 10 |
| d_model / heads / FFN | 1024 / 16 / 2816 |
| Vocabulary | 48,000 (SentencePiece BPE, shared) |
| Max sequence length | 512 tokens |
| Positional encoding | sinusoidal |
| Language conditioning | learned language embeddings added to encoder and decoder inputs |
| Tied embeddings | input embedding = output projection |
| Training steps | 5,930,000 |
Evaluation — FLORES-200 devtest
1012 sentences per direction, beam size 5, sacreBLEU.
| Direction | BLEU | spBLEU | chrF |
|---|---|---|---|
| it → en | 24.79 | 28.80 | 56.39 |
| en → it | 20.95 | 26.26 | 52.49 |
| fr → en | 29.18 | 32.09 | 57.26 |
| en → fr | 26.33 | 29.58 | 54.70 |
| es → en | 18.88 | 22.09 | 50.75 |
| en → es | 17.00 | 19.89 | 47.05 |
| pt → en | 33.34 | 36.42 | 60.55 |
| en → pt | 27.84 | 31.45 | 56.72 |
| de → en | 24.34 | 26.80 | 52.58 |
| en → de | 15.88 | 19.10 | 46.68 |
| nl → en | 18.60 | 21.36 | 48.09 |
| en → nl | 14.01 | 17.66 | 45.00 |
| ro → en | 26.95 | 29.92 | 56.88 |
| en → ro | 18.55 | 22.73 | 49.15 |
| pl → en | 15.46 | 18.28 | 45.94 |
| en → pl | 8.69 | 14.45 | 38.95 |
| cs → en | 22.03 | 24.61 | 51.64 |
| en → cs | 13.50 | 18.33 | 42.38 |
| sv → en | 29.98 | 31.76 | 56.83 |
| en → sv | 23.74 | 26.52 | 52.72 |
| Average X → en | 24.36 | 27.21 | 53.69 |
| Average en → X | 18.65 | 22.60 | 48.58 |
| Average (all 20) | 21.50 | 24.91 | 51.14 |
Usage
pip install torch sentencepiece safetensors
python inference.py --src it --tgt en "Domani mattina devo passare in farmacia."
python inference.py --src en --tgt de --beam 5 "Please back up your data before updating."
from inference import load, translate
model, sp, cfg, device = load()
print(translate("Il treno è in ritardo.", "it", "fr", model, sp, cfg, device))
Training data
- FineTranslations (HuggingFaceFW/finetranslations) — web documents paired with English translations, all 10 languages
- Europarl, Tatoeba
- Italian YouTube subtitles translated into 10 languages with NLLB-600M (distilled)
- Small synthetic sets targeting specific weaknesses (proper names, colloquial phrases)
Every pair is used in both directions during training.
Limitations
- Idioms and colloquial expressions are often translated literally ("stanco morto" → "tired of death").
- Domain jargon (gaming/streaming, some IT terms) is inconsistent.
- Polish and Czech are the weakest languages, especially en → pl / en → cs; basic vocabulary errors still occur.
- Inputs longer than ~512 tokens are truncated: split long texts into sentences.
- Not suitable for high-stakes use (medical, legal) without human review.
Citation
@misc{iris300m,
author = {Michelangelo Di Nicola, Roberto D'Arcangelo},
title = {Iris-300M: a compact multilingual European translation model},
year = {2026},
publisher = {Hugging Face}
}