harrier-oss-v1

harrier-oss-v1 is an open-source family of multilingual text-embedding models from Microsoft that delivers strong retrieval and semantic-understanding quality across 94 languages. The models use a decoder-only architecture with last-token pooling and L2 normalization, and they accept natural-language instructions so that the same checkpoint can be specialized for retrieval, classification, clustering, or similarity tasks. Three sizes — 270M, 0.6B, and 27B parameters — are released, covering the full latency-versus-accuracy spectrum from on-device retrieval to high-throughput server workloads.

harrier-oss-v1 addresses the production need for embeddings that combine wide language coverage with deployment flexibility, eliminating the common pattern of stitching together language-specific or task-specific models. Open weights at multiple scales let teams pick a configuration matched to their latency and cost constraints while preserving consistent semantic quality across the world’s languages. The release is part of Microsoft’s broader effort to put production-grade open models — alongside Phi and BitNet — into developers’ hands so that high-quality multilingual retrieval is no longer the exclusive territory of closed APIs.