Microsoft open-sources three versions of the Harrier text embedding model, with the 27B version topping the multilingual MTEB v2 leaderboard.

BlockBeatNews · 2026-03-30T11:07:06+00:00

Microsoft has open-sourced the multilingual text embedding model named harrier-oss-v1 on Hugging Face, available in 270M, 0.6B, and 27B versions. It supports various tasks and performs excellently on the Multilingual MTEB v2 benchmark. The 27B version topped the leaderboard on its release day.

BlockBeatNews

2026-03-30 11:07:06

Abstract generation in progress

According to 1M AI News monitoring, Microsoft has open-sourced a multilingual text embedding model family on Hugging Face called harrier-oss-v1, with three tiers: 270M, 0.6B, and 27B. The model card shows that this series uses a decoder-only architecture, last-token pooling, and L2 normalization, supports up to 32768 tokens, and can be used for retrieval, clustering, semantic similarity, classification, bilingual mining, and re-ranking.

Multilingual MTEB v2 is an industry-standard multilingual text embedding benchmark, primarily testing tasks such as retrieval, classification, clustering, and semantic similarity. Microsoft’s model card states that the three tiers achieve scores of 66.5, 69.0, and 74.3 on this benchmark, respectively, and that the 27B version topped the leaderboard on the day it was released. The 270M and 0.6B versions also additionally use larger embedding models for knowledge distillation. All three models are released under the MIT license.

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.