EuroLLM-1.7B-Instruct: EU Multilingual LLM

• EuroLLM-1.7B-Instruct is an open-weight, instruction-tuned multilingual LLM designed for high-quality European language translation and real-time XR deployments.
• The model employs a 24-layer decoder-only Transformer with rotary positional embeddings and SwiGLU activations, ensuring efficient inference and memory use.
• Benchmark evaluations show significant gains in BLEU and COMET scores, outperforming larger models like NLLB-200 in translation speed and accuracy.

EuroLLM-1.7B-Instruct is an open-weight, instruction-tuned multilingual causal LLM specifically designed to deliver high-quality natural language understanding and generation across all official European Union languages and several additional languages. Built upon a 1.7-billion-parameter Transformer architecture, this compact model demonstrates state-of-the-art performance in European language translation, achieving higher BLEU and COMET scores than comparably sized and even substantially larger models. Its primary deployment context to date has been in modular, real-time AI systems for extended reality (XR) applications, where it outperforms Meta's NLLB-200 and other competitive baselines in both translation quality and operational efficiency (Tantaroudas et al., 7 Apr 2026, Martins et al., 2024).

1. Model Architecture and Specification

EuroLLM-1.7B-Instruct is based on a 24-layer, decoder-only Transformer employing pre-layer normalization and rotary positional embeddings. Its key architectural features include:

• Hidden dimension: $d_\text{model} = 2048$
• Feed-forward network dimension: $4 \times d_\text{model} \times 1.375 \approx 5632$ (utilizing the SwiGLU activation with a 1.375 expansion factor)
• Attention configuration: 16 attention heads for queries, key/value grouped-query attention (GQA) with 8 key/value heads to accelerate inference
• Normalization and activations: RMSNorm in a pre-norm arrangement, SwiGLU activations
• Positional encoding: RoPE with base $\theta=10000$
• Sequence length: Up to 4096 tokens

Parameter distribution is as follows:

• Embedding matrix: 0.262B
• LM head: 0.262B
• Transformer (non-embedding) parameters: 1.133B
• Total: 1.657B (rounded to 1.7B)

This configuration results in a highly efficient model for both memory and inference speed, with model load times of approximately 41 s in float16 precision and a 3.5 GB VRAM footprint on an NVIDIA RTX 4060 (Tantaroudas et al., 7 Apr 2026, Martins et al., 2024).

2. Multilingual Tokenization and Pretraining Data

The tokenizer is a byte-pair encoding (BPE) model with byte-fallback, trained on a corpus of approximately 4 trillion tokens spanning all target languages. The vocabulary comprises 128,000 tokens, selected to minimize subword fertility across diverse writing systems—Latin, Cyrillic, Greek, Arabic, Han, Devanagari, Hangul, Kana—while controlling parameter growth. Empirical fertility analysis shows much lower pieces-per-word compared to 32K-token vocabularies and near parity with 256K-token setups for European languages (Martins et al., 2024).

Pretraining draws on a balanced mixture:

• Web data: English (FineWeb-edu), high-resource EU languages (RedPajama-Data-v2), other European and global languages (HPLT, MADLAD-400, CulturaX, mC4)
• Parallel data: Public WMT/EU xx↔en, filtered by Bicleaner and CometKiwi-22 scores
• Code/math: The Stack, Algebraic-stack, Open-web-math
• High-quality text: Wikipedia, ArXiv, Books, Apollo

Mixture ratios start with English at 50%, with remaining languages at 45% and code/math at 5%, shifting in the later annealing phase to prioritize high-quality content and decrease the English fraction (Martins et al., 2024).

3. Instruction Tuning Methodology

Instruction tuning, which distinguishes the Instruct variant from the Base model, employs the "EuroBlocks" dataset of approximately 1 million prompt–response pairs across all supported languages. Sources include:

• Human-written and synthetic dialogs: OpenHermes-2.5, Aya
• High-quality machine translation instances: NTREX-128, FLORES-200-dev, WMT-21, WMT-22

The fine-tuning objective is a masked cross-entropy loss focusing solely on target tokens. Training is executed in bfloat16 mixed precision with sequence packing and a constant learning rate of $7 \times 10^{-6}$ over 4 epochs (approximately 6,000 update steps). Control tokens <|im_start|>/{system,user,assistant} and <|im_end|> are reserved for structure (Martins et al., 2024).

Instruction tuning aligns generation behavior with human-like translation prompts, enabling robust zero-shot generalization for European languages. It significantly enhances BLEU, COMET, and throughput metrics relative to both the Base variant and competitive translation models (Tantaroudas et al., 7 Apr 2026).

4. Evaluation: Benchmarks and Comparative Performance

EuroLLM-1.7B-Instruct has been evaluated on established multilingual general knowledge, reasoning, and translation benchmarks:

General Knowledge and Reasoning

• HellaSwag (10-shot): 0.475 avg accuracy (vs Gemma-2b 0.460, TinyLlama 0.430)
• ARC Challenge (25-shot): 0.327 (slightly behind Gemma-2b at 0.340) (Martins et al., 2024)
• Instruction tuning provides an absolute gain of 0.005–0.010 over the base model.

Machine Translation (Zero-Shot, Greedy Decoding, COMET-22)

• FLORES-200 dev (EN→XX): 86.8 (Gemma-2b-Instruct 84.0, Gemma-7b-Instruct 87.2)
• FLORES-200 dev (XX→EN): 87.1 (Gemma-2b 83.5, Gemma-7b 87.5)
• WMT-23: 83.1 (Gemma-2b 80.0, Gemma-7b 83.5)
• WMT-24: 79.4 (Gemma-2b 76.6, Gemma-7b 79.8) (Martins et al., 2024)

Real-World XR Translation Pipeline (English→French)

Model	BLEU	Latency (s)	VRAM (GB)	Load Time (s)
NLLB-200-distilled-600M	79.25	0.596	≈2.5	26.63
EuroLLM-1.7B Base	27.58	1.509	≈3.5	25.37
EuroLLM-1.7B Instruct	84.34	0.529	≈3.5	40.99

EuroLLM-1.7B-Instruct outperforms NLLB-200 by over 5 BLEU points and is ≈10% faster in inference, despite a higher model-load overhead. The Base variant fails to achieve competitive translation performance in the absence of instruction tuning (Tantaroudas et al., 7 Apr 2026).

5. System Integration and Deployment Context

Within modular AI architectures for XR—specifically, Unity/Meta Quest 3-based environments—EuroLLM-1.7B-Instruct operates as a plug-in microservice, accepting RESTful translation requests immediately downstream from automatic speech recognition (via OpenAI Whisper) and feeding text to AWS Polly for speech synthesis and to IS rendering modules for sign-language representation. Docker-based deployment within AWS cloud infrastructure enables scalable, real-time translation with average response times below 0.53 s (Tantaroudas et al., 7 Apr 2026).

The model's compact size and latency profile allow it to serve European language translation within latency constraints suitable for immersive, interactive learning environments. A plausible implication is suitability for other cross-modal or context-constrained applications where translation and NLU are required.

6. Design Rationale, Scaling Laws, and Future Directions

Empirical scaling law analysis was conducted for EuroLLM, modeling loss as a function of non-embedding parameter count and language data availability. Marginal translation gains are observed when increasing the share of parallel data up to ~20%, after which return diminishes. Repeating high-quality language resources improves held-out Wikipedia performance with no adverse effect on web-text capabilities (Martins et al., 2024).

The Instruct variant’s performance advantage is attributed to:

1. Instruction Tuning: Strongly aligning the generation objective with translation tasks and prompt structures, evidenced by a leap from BLEU 27.58 (Base) to 84.34 (Instruct) (Tantaroudas et al., 7 Apr 2026).
2. Architectural Efficiency: The causal stack with parallel attention enables faster inference and effective prompt use, outperforming sparsely gated MoE models like NLLB-200 for medium-sized language pairs.

Limitations include suboptimal general reasoning at 1.7B scale, weaker performance for very low-resource languages, and lack of reinforcement learning or human-feedback-based alignment. Plans for future work include scaling to larger parameter counts, further filtering and expanding data coverage, studying longer-context capabilities, and integrating RLHF or advanced alignment mechanisms (Martins et al., 2024).

7. Summary and Significance

EuroLLM-1.7B-Instruct represents a rigorous effort to produce a high-coverage, instruction-aligned, open-weight LLM optimized for European language tasks. It achieves near parity with much larger models on machine translation, surpasses established baselines (including NLLB-200) in both quality and inference speed, and integrates seamlessly into demanding, latency-sensitive, multilingual XR deployments (Tantaroudas et al., 7 Apr 2026, Martins et al., 2024). Its empirical design, open research methodology, and clear benchmarking position it as a reference LLM for research and practical applications within European and multilingual NLP settings.