Meta's NLLB-200-distilled-1.3B with its 256k-token, 200-language vocabulary pruned to 39,212 tokens — every Hebrew-script piece, frequent-English pieces, character-level fallbacks for the served scripts, and all language codes — then quantised to int8 ONNX for running in the browser with transformers.js.

No weights were retrained. Pruning removes embedding rows the language pair cannot reach; on a 12-segment Yiddish benchmark against a human-reviewed translation the pruned model's output is character-identical to the full model (chrF 49.9 both), and the int8 conversion costs 0.4 chrF (49.5). The distilled-600M scores 35.3 on the same harness — this build exists because the quality gap is that large.

Translation output is still a draft: the model follows sentence shape well and can get concrete nouns wrong. Keep the source text beside it.

The underlying model is Meta AI's NLLB-200 (CC-BY-NC-4.0). This repository is a format conversion with vocabulary pruning and carries the same non-commercial licence. All credit for the translation capability belongs to the NLLB team.