translation
HF group: NLP · Status: 🟡 partial (works through chat LLM)
What it is
Text in language A → text in language B. Specialized translation models meaningfully outperform chat LLMs on low-resource languages.
Open-weight models
| Model | Params | Released | License | Quality | Notes |
|---|---|---|---|---|---|
| MADLAD-400 (3B / 7B / 10B) | 3-10 B | 2023 | Apache-2.0 | 419 languages | Apache-2.0 — best for commercial. |
| NLLB-200 | 600 M – 54 B | 2022 | CC-BY-NC | 200 languages | Non-commercial license. |
| OPUS-MT (per-pair) | ~80 M | 2020+ | Apache-2.0 | Solid for major pairs | Many small per-pair models. |
| Meta Omnilingual MT (OMT) | varies | 2026-03 | TBD | 1,600 languages | Newest; license details still being clarified. |
| Any chat LLM | varies | various | various | Strong on major pairs, weak on rare | Convenient but lossy on low-resource. |
Infrastructure required
Inference
- 🟡 Works today via chat LLM in
locara-llama. - ❌ Encoder-decoder runtime (BART/MBART class) for specialist models. llama.cpp doesn’t support these natively; would need Candle or ONNX Runtime path.
Input
- Plain text + source/target language codes.
Output
- Plain text.
Storage
- ✅ Weights cache.
- App-side: optionally cache translations.
Interaction (IPC + SDK)
- 🟡 Today:
llm.chatwith a “translate to X” prompt. - ❌ Specialized:
translate.text({ text, from, to })IPC.
Capabilities (manifest)
capabilities.models[]for the translator (or chat LLM fallback).
Gaps
- Apache-2.0 specialized model (MADLAD-400) for quality on low-resource languages.
- Encoder-decoder runtime — shared need with summarization specialists and Donut for DocVQA.
See also
text-to-text— fallback todaysummarization— same encoder-decoder need- Index:
../modalities-and-models-survey.md