Microsoft's Bing team has open-sourced its latest word embedding model, "Harrier." The model excels in multilingual MTEB v2 benchmarks, supports over 100 languages, and offers enhanced language processing capabilities. Harrier is trained on over 2 billion examples and synthetic data from GPT-5, utilizing a 32,000-token context window for improved accuracy in multilingual tasks. It comes in three versions: a full 2.7-billion-parameter model, alongside smaller 60-million and 270-million-parameter variants for low-performance hardware. All models are available on Hugging Face under an MIT license. Embedding models are crucial for AI systems, particularly in search, information retrieval, and data organization. Harrier's release aims to advance AI applications by enabling more complex, multi-step tasks. Microsoft plans to integrate Harrier into Bing and next-generation AI agent services, boosting the search engine's competitiveness and meeting demands for efficient information processing.