Microsoft Bing Team Open-Sources "Harrier" Embedding Model

2026-04-08

Microsoft's Bing team has unveiled "Harrier," a new open-source embedding model designed to enhance AI capabilities. Supporting over 100 languages, Harrier features a 32,000-token context window and was trained on a dataset comprising more than 2 billion examples alongside synthetic data generated by GPT-5. According to the development team, Harrier currently tops the multilingual MTEB v2 benchmarks, outperforming proprietary models from both OpenAI and Amazon.

In addition to the primary 27-billion-parameter model, the release includes two smaller variants with 0.6 billion and 270 million parameters, optimized for hardware with lower performance capabilities. All three models are accessible on Hugging Face under the MIT license. Looking ahead, the team plans to integrate this technology directly into Bing and utilize it within new AI agent grounding services.

Embedding models are responsible for searching, retrieving, and organizing the information required by AI systems to deliver accurate responses. Microsoft notes that these models are becoming increasingly critical as AI agents independently undertake more complex, multi-step tasks.

RECENT AI NEWS

RECENT AI TOOLS