Swan-Small tops benchmarks for Arabic dialect embeddings

The quest for a "Small Multilingual Model" (SMM) highlighting the Ar-Fr-En triad exposes the persistent "script gap" in general-purpose embeddings. While standard models handle formal text well, casual code-switching and transliteration require specialized architectural choices like Swan-Small. This model outperforms larger general-purpose bases on dialectal benchmarks, while specific needs may still favor dialect-focused models like DziriBERT or TunBERT. For interactive gaming and edge applications where low latency is paramount, these compact, dialect-aware embeddings provide the necessary semantic precision without the overhead of massive LLMs.