Harrier

Open-source multilingual text embedding model family optimized for agentic AI systems. Decoder-only architecture with last-token pooling and L2 normalization. Trained via contrastive learning on large-scale multilingual data with knowledge distillation from larger embedding models. Supports 94 languages with 32K token context.

The 27B variant achieves #1 on MTEB-v2 (74.3), the industry-standard multilingual embedding benchmark. Designed as a foundational layer for memory, ranking, and orchestration in agent-based systems — enabling cross-source retrieval, persistent memory over extended interactions, and dynamic context updates across multi-step tasks. MIT License.