Meta's Omnilingual MT covers 1,600 languages, 1B–8B models rival 70B LLMs
kalyan_kpl · x · 2026-10-28
Meta FAIR introduces Omnilingual MT, the most broadly benchmarked MT system to date covering over 1,600 languages. It explores two architectures: OMT-LLAMA (1B/3B/8B decoder-only models built on Llama3 with multilingual continual pretraining) and OMT-NLLB (a 3B encoder-decoder with a novel three-stage strategy exploiting non-parallel data via autoencoding). The specialized 1B–8B models match or exceed 70B general-purpose LLMs, and human annotators rate translation as passable for 440+ languages—double the prior state of the art of 200—with zero-shot translation into 1,200 more.
More from Models
- Open model Flint hits #2 trending on Hugging Face, team outlines next steps — josh_tobin_ · 2026-10-28
- YacineMTB rails: even the best LLMs are 'complete dogshit' and he's tired of it — yacineMTB · 2026-10-28
- Omni-InteractScope benchmarks proactive interaction in continuous omnimodal streams — Qwen · 2026-10-28
- Mistral releases Timestral time-series foundation models, tops GIFT-Eval — mistralai · 2026-10-28
- Gemini 4 Argon Hits 1.73% WER on OCR Benchmark, Surpassing Human Experts — gabriberton · 2026-10-28
- Uncensored LFM3-1.7B GGUF weights released on Hugging Face — kalyan_kpl · 2026-10-28