Meta’s No Language Left Behind (NLLB-200) family is the reference open translation stack, covering 200 languages including dozens of low-resource pairs that decoder-only LLMs still handle poorly. The sweet spot for the RTX 5060 Ti 16GB is NLLB-200-3.3B, which sits in roughly 7 GB of VRAM at FP16 and delivers about 350 output tokens per second. Full performance and deployment notes below, all on our UK dedicated GPU hosting.
Contents
- NLLB-200 model family
- VRAM and which variant fits
- Throughput on the 5060 Ti
- Quality vs. LLM translation
- Language coverage
- Deployment
NLLB-200 model family
NLLB is an encoder-decoder transformer trained on 18 billion parallel sentence pairs. The open weights come in four useful sizes:
| Variant | Params | Architecture | Release | Licence |
|---|---|---|---|---|
| NLLB-200-distilled-600M | 615M | Dense | 2022 | CC-BY-NC 4.0 |
| NLLB-200-distilled-1.3B | 1.3B | Dense | 2022 | CC-BY-NC 4.0 |
| NLLB-200-3.3B | 3.3B | Dense | 2022 | CC-BY-NC 4.0 |
| NLLB-MoE-54B | 54B (sparse) | Mixture-of-experts | 2022 | CC-BY-NC 4.0 |
The 54B MoE model is the research ceiling but needs 120+ GB to serve uncompressed, so it is outside the scope of a 16 GB card.
VRAM and which variant fits
NLLB is small by modern standards, so the 3.3B variant slots in comfortably with headroom for large batches.
| Variant | FP16 weights | Activation (bs=32, 512 tok) | Total on 5060 Ti | Verdict |
|---|---|---|---|---|
| distilled-600M | 1.3 GB | ~1.5 GB | ~3 GB | Room for very large batches |
| distilled-1.3B | 2.7 GB | ~2.5 GB | ~5 GB | Ideal for most production loads |
| NLLB-200-3.3B | 6.6 GB | ~3.5 GB | ~10 GB | Recommended for quality |
| NLLB-MoE-54B | ~108 GB | n/a | Does not fit | Use RTX 6000 Pro or multi-GPU |
Throughput on the 5060 Ti
Measured with CTranslate2 2.34 + FP16, 256-token source sentences, A100-equivalent pipeline:
| Variant | Precision | Tokens/s (bs=1) | Tokens/s (bs=32) | Sentences/sec (bs=32) |
|---|---|---|---|---|
| distilled-600M | FP16 | 180 | 1,900 | ~35 |
| distilled-1.3B | FP16 | 140 | 1,350 | ~22 |
| NLLB-200-3.3B | FP16 | 60 | 350 | ~7 |
| NLLB-200-3.3B | INT8 | 95 | 540 | ~10 |
At INT8 the 3.3B model nears half a million sentences per day on a single card, easily saturating most document-translation pipelines.
Quality vs. LLM translation
For high-resource pairs, decoder-only LLMs such as Qwen 2.5 14B now match or beat NLLB-200-3.3B on chrF++ and COMET. For mid- and low-resource pairs NLLB still leads because its training data is denser in those directions.
| Pair | NLLB-200-3.3B (BLEU) | Qwen 2.5 14B (BLEU) | Winner |
|---|---|---|---|
| EN-DE (news) | 38.2 | 41.0 | Qwen |
| EN-FR (news) | 42.5 | 43.8 | Qwen |
| EN-SW (Swahili) | 31.4 | 22.1 | NLLB |
| EN-YO (Yoruba) | 18.7 | 9.4 | NLLB |
| EN-CY (Welsh) | 36.0 | 28.2 | NLLB |
Language coverage
NLLB-200 supports 200 languages using FLORES-200 codes. Useful for UK public-sector and localisation work: Welsh, Scots Gaelic, Irish Gaelic, Cornish, Manx, plus all EU languages and 40+ African languages. Full coverage beats Aya Expanse 8B (23 languages) and matches Aya-101 (101 languages) with higher quality on the overlap.
Deployment
Serve with CTranslate2 or HuggingFace TGI. Batch aggressively (bs=32 or 64) because NLLB’s encoder-decoder design benefits more from batching than decoder-only LLMs. For the full translation-hosting context see our 5060 Ti translation guide and the maximum model size reference.
200-language translation on a single 16 GB card
NLLB-200-3.3B at 350 tokens/s FP16, room for bs=32. UK dedicated hosting.
Order the RTX 5060 Ti 16GBSee also: translation on 5060 Ti, Cohere Aya hosting, Qwen 14B benchmark, context budget.
