Microsoft Unveils Harrier, a New Open-Source Embeddings Model for the Agentic Web

Microsoft Bing team has announced an open-sourced a new embedding model, Harrier, designed to improve grounding for AI systems as they move from answering questions to taking action. The announcement comes just days after Microsoft released new models, including MAI-Image-2 and MAI-Transcribe-1, which offers industry leading speech-to-text accuracy.

Speaking of Harrier, the Bing team notes that the model addresses one of the toughest challenges for AI: grounding. By connecting information across diverse sources into coherent, meaningful responses, embeddings like Harrier help agents deliver accurate answers faster and with fewer errors. Microsoft says better embeddings translate into higher factual accuracy, lower latency, and more consistent agent behavior over multi-step tasks.

Designed for the “agent era,” Harrier supports over 100 languages, offers a 32k token context window, and produces fixed-size embeddings for seamless integration with vector search systems. According to Microsoft, it ranks first on the multilingual MTEB-v2 benchmark as of April 6, 2026, outperforming both open-source and proprietary alternatives.

Harrier was created by Microsoft using GPT-5, as it helped henerate synthetic multilingual data, resulting in over 2 billion weakly-supervised examples and 10 million high-quality fine-tuning samples. The company also applied knowledge distillation from larger teacher models to boost performance of smaller models, ensuring deployment flexibility on a range of devices. Worth noting that Harrier-OSS-v1-27B leads the benchmark, while smaller versions like Harrier-OSS-v1-0.6b and Harrier-OSS-v1-270M target low-end devices without sacrificing accuracy.

Do note that Harrier is part of a broader effort to improve AI grounding at scale, with plans to integrate the technology into Bing and future agentic services. Microsoft emphasizes that grounding, knowing what to use, when, and in what context, remains the foundation for user trust in AI agents.