Back to Gap Map Model components / Embeddings & retrieval models

Harrier-OSS

open weights / Overall score: 4.4(strong)

Harrier-OSS is Microsoft's open-weight multilingual text embedding family, released by the Bing team in three sizes: 270M, 0.6B and 27B parameters. The models are decoder-only encoders with last-token pooling and L2 normalization, support more than 100 languages and a 32,768-token context, and take a one-sentence instruction on the query side. They target retrieval, clustering, semantic similarity, classification, bitext mining and reranking, and the 27B checkpoint scores 74.3 on the Multilingual MTEB v2 benchmark.

Openness

3 high confidence

3.0

weights: open(three ungated MIT safetensors checkpoints)
data: closed(over 2B weakly-supervised pairs plus GPT-5 synthetic data, described but not released)
code: partial(inference usage and MTEB evaluation prompts only)
license: MIT(OSI)

Harrier-OSS ships ungated weights under MIT, and Microsoft describes the training pipeline and synthetic-data generation step in prose, but neither the corpus nor the code that produced it is released. That keeps it at open weights rather than a fully open release.

https://blogs.bing.com/search/April-2026/Microsoft-Open-Sources-Industry-Leading-Embedding-Model recorded 2026-09-11
Microsoft describes a data pipeline producing more than 2 billion weakly-supervised examples and over 10 million high-quality fine-tuning examples, with GPT-5 used to synthesize multilingual pairs; none of it is offered for download.
https://huggingface.co/api/models/microsoft/harrier-oss-v1-27b recorded 2026-09-11
gated: false; private: false; disabled: false; cardData license mit; sharded safetensors weight files present.
https://huggingface.co/microsoft/harrier-oss-v1-27b/raw/main/README.md recorded 2026-09-11
license: mit in the card front matter; three downloadable checkpoints listed; training described as contrastive learning on a large-scale multilingual mixture with distillation, naming no dataset and no training code; usage snippets and an evaluation-prompt file only.

Adoption

3 high confidence

3.0

Adoption is measured as Hugging Face downloads summed across the three shipped checkpoints, most of them for the small, on-device 270M and 0.6B sizes. The 27B flagship that tops the benchmark needs serious hardware and is downloaded far less often.

https://huggingface.co/api/models/microsoft/harrier-oss-v1-0.6b recorded 2026-09-11
downloads: 243108; gated: false; private: false.
https://huggingface.co/api/models/microsoft/harrier-oss-v1-270m recorded 2026-09-11
downloads: 515530; gated: false; private: false.
https://huggingface.co/api/models/microsoft/harrier-oss-v1-27b recorded 2026-09-11
downloads: 6909; gated: false; private: false.

Capability

5 high confidence

5.0

Harrier-OSS-v1-27B reports the strongest Multilingual MTEB v2 result claimed in this category -- 74.3, across more than 100 languages at a 32k context -- and every other product here is measured against that figure. It is Microsoft's own claim rather than an independent leaderboard reading, since the card and the announcement are the same publisher.

https://blogs.bing.com/search/April-2026/Microsoft-Open-Sources-Industry-Leading-Embedding-Model recorded 2026-09-11
Harrier is described as ranking 1st on the multilingual MTEB-v2 benchmark as of April 6, 2026, supporting more than 100 languages with a 32k context window.
https://huggingface.co/microsoft/harrier-oss-v1-27b/raw/main/README.md recorded 2026-09-11
Model table gives MTEB v2 scores of 66.5 (270m), 69.0 (0.6b) and 74.3 (27b) at 32,768 max tokens each, and states the models achieve state-of-the-art results on Multilingual MTEB v2 as of the release date.

Verified 2026-09-11