Microsoft AI Releases Harrier-OSS-v1: New Multilingual Embedding Family Achieves SOTA on MTEB v2

Microsoft released Harrier-OSS-v1, a family of three multilingual embedding models.

The models come in 270M, 0.6B and 27B sizes and all support a 32,768-token context window.

Harrier-OSS-v1 uses decoder-only architectures with last-token pooling and achieved SOTA on Multilingual MTEB v2.

The smaller models were improved via knowledge distillation and the family is instruction-tuned for query-side prompts.