YT · YOUTUBE// 10d agoMODEL RELEASE

Microsoft drops Harrier-OSS-v1 multilingual embeddings

Microsoft released Harrier-OSS-v1, a family of decoder-only multilingual text embedding models supporting 94 languages and 32k context. The models achieve SOTA results on Multilingual MTEB v2 through architectural shifts and knowledge distillation.

// ANALYSIS

Microsoft is aggressively moving into the open-weights embedding space, proving that decoder-only backbones are the future for retrieval.

–32k token context window eliminates complex chunking strategies and "lost in the middle" issues for long-form RAG.
–Smaller variants (270M, 0.6B) punch far above their weight due to distillation from 27B+ parameter backbones.
–94-language support and instruction-tuning make it a versatile foundation for global retrieval and classification tasks.

// TAGS

harrier-oss-v1llmembeddingmultilingualopen-sourcerag

DISCOVERED

10d ago

2026-04-02

PUBLISHED

10d ago

2026-04-02

RELEVANCE

9/ 10

AUTHOR

AI Revolution