Microsoft's Bing Team Open-Sources 'Harrier' — a 27B Embedding Model That Tops the Multilingual MTEB Benchmark

Microsoft's Bing team has open-sourced Harrier, a multilingual embedding model trained on over two billion examples plus synthetic data generated from GPT-5. The model achieves state-of-the-art performance on the multilingual MTEB v2 benchmark, outperforming proprietary embedding models from OpenAI and Amazon, and is available in three sizes on Hugging Face under the MIT license — one of the most permissive open-source licenses in AI model releases.

What Harrier Does

Embedding models convert text into numerical vectors that capture semantic meaning, enabling AI systems to search, retrieve, compare, and organize information by conceptual similarity rather than exact keyword match. They are foundational infrastructure for retrieval-augmented generation (RAG) systems, semantic search, document clustering, and the "grounding" mechanisms that help AI agents access relevant context before generating responses. As agentic AI systems take on more complex, multi-step tasks — autonomously browsing documents, querying knowledge bases, and synthesizing information across sources — the quality of the underlying embedding model directly affects how accurately the agent retrieves what it needs.

Harrier's 32,000-token context window is particularly significant for enterprise use cases. Most embedding models have context windows in the 512–8,192 token range, which means long documents must be chunked before embedding — a process that can lose context across chunk boundaries and degrade retrieval quality for content that requires understanding document-level structure. A 32K context window allows Harrier to embed much longer document segments as coherent units, improving retrieval precision for legal documents, technical reports, and research papers.

Why Open-Source MIT Licensing Matters

The MIT license imposes essentially no restrictions on commercial use, modification, or redistribution. This is a more permissive release than Meta's Llama licenses (which restrict certain commercial deployments) and contrasts with the typical approach of proprietary API-only embedding models from OpenAI and Cohere. MIT licensing means any company can deploy Harrier internally, fine-tune it on proprietary data, or build commercial products on top of it without royalties or usage restrictions — including at scales that would be economically prohibitive through a per-token API.

Integration Plans

Microsoft plans to integrate Harrier into Bing search and into new "grounding services" for AI agents — the infrastructure that connects agents to external knowledge sources. Making the model's underlying architecture publicly available while deploying it in production at Bing scale gives Microsoft a feedback mechanism between research and deployment that benefits both the open-source community and the company's own product roadmap. The two smaller variants (0.6B and 270M) are designed for edge and on-device scenarios where the full 27B model is impractical, making Harrier a complete model family rather than a single release.