Microsoft Just Released Harrier-OSS-v1: A New Family of Multilingual Embedding Models

Description

Microsoft’s new Harrier-OSS-v1 suite is a targeted play for the retrieval layer, offering a trio of multilingual embedding models (270M, 0.6B, and 27B) that hits SOTA status on the Multilingual MTEB v2 benchmark at release. Ditching traditional BERT-style encoders for decoder-only architectures—these models utilize last-token pooling and a massive 32,768-token context window to handle long-form data without the typical "chunking" penalty. While the 27B giant provides high-dimensional precision (5,376 dims), the smaller variants benefit from knowledge distillation to maintain high representation quality at a lower compute cost..... Full analysis: https://www.marktechpost.com/2026/03/30/microsoft-ai-releases-harrier-oss-v1-a-new-family-of-multilingual-embedding-models-hitting-sota-on-multilingual-mteb-v2/ Model Weights: https://huggingface.co/microsoft/harrier-oss-v1-270m @Microsoft @MicrosoftAzure @MicrosoftResearch #artificialintelligence #llm #ai #opensource #llms #transformers