Adapting Multilingual Embedding Models to Turkish via Cross-Lingual Tokenizer Surgery and Offline Distillation

Researchers released embeddingmagibu-200m, a 200M-parameter Turkish sentence embedding model adapted from a multilingual teacher.

It uses a Turkish-optimized 131k tokenizer and offline distillation from precomputed embeddings, cutting training cost.

The model matches or beats the teacher on major benchmarks, including 77.55 Pearson and 77.45 Spearman on STSbTR and a 63.9% mean score on TR-MTEB.

The team also released the model, tokenizer, datasets, and tooling, making it easier to build lower-cost Turkish NLP systems.