Table of Contents
NLLB-200 (Meta) is purpose-built for translation across 200 languages. Three sizes — 600M, 1.3B, 3.3B. All fit comfortably on a 5060 Ti.
TL;DR
NLLB-200 3.3B at FP16 on a 5060 Ti hits ~3,000 translation tokens/sec. Smaller variants 5,500+ tok/s. Enough for tens of thousands of pages/day on a single £119/mo server.
NLLB variants
| Variant | Params | VRAM | Quality | Speed |
|---|---|---|---|---|
| NLLB-200 distilled 600M | 600M | 1.2 GB | Good | Fastest |
| NLLB-200 distilled 1.3B | 1.3B | 2.6 GB | Strong | Fast |
| NLLB-200 3.3B (full) | 3.3B | 7 GB | Strongest | Solid |
Throughput
| Model | Translation tok/s |
|---|---|
| NLLB-200 600M | ~10,000 |
| NLLB-200 1.3B | ~5,500 |
| NLLB-200 3.3B | ~3,000 |
Verdict
NLLB-200 on a 5060 Ti is the cheapest credible 200-language translation deployment. For specific high-resource pairs, an LLM-based translator (Qwen 2.5 7B) gives slightly better quality at lower throughput.
Bottom line
Self-hosted translation on a 5060 Ti at £119/mo replaces DeepL Pro / Google Translate API at meaningful volume. See 5060 Ti for translation.
