model releaseMicrosoft

Microsoft releases Harrier embedding models with 32K context window, achieving 74.3 on MTEB v2

TL;DR

Microsoft released the Harrier-OSS embedding model family, comprising three variants with 270M, 600M, and 27B parameters. The largest model achieves 74.3 on the Multilingual MTEB v2 benchmark. All models support 32,768 max tokens and multilingual inputs across 40+ languages.

2 min read
0

Microsoft Releases Harrier Embedding Models with State-of-the-Art Multilingual Performance

Microsoft has released Harrier-OSS, a family of multilingual text embedding models designed for retrieval, clustering, semantic similarity, classification, and reranking tasks. The open-source models are available on Hugging Face.

Model Specifications

The Harrier family includes three variants:

Model Parameters Embedding Dimension Max Context MTEB v2 Score
harrier-oss-v1-270m 270M 640 32,768 tokens 66.5
harrier-oss-v1-0.6b 600M 1,024 32,768 tokens 69.0
harrier-oss-v1-27b 27B 5,376 32,768 tokens 74.3

All models use decoder-only architectures with last-token pooling and L2 normalization to generate dense embeddings. The 270M and 600M variants employ knowledge distillation from larger embedding models during training.

Training and Capabilities

Microsoft trained all variants using contrastive learning on multilingual datasets covering diverse embedding tasks. The models support 40+ languages including English, Spanish, French, German, Chinese, Japanese, Arabic, and Hindi.

Key capabilities span:

  • Dense passage retrieval
  • Semantic similarity scoring
  • Text clustering
  • Bitext mining
  • Zero-shot classification and reranking

Each model requires task-specific instructions appended to queries during inference—for example, "Instruct: Retrieve semantically similar text\nQuery: [user query]". Documents do not require instructions.

Technical Details

The models are compatible with both Sentence Transformers and native Hugging Face Transformers libraries. They use BF16 tensor precision and are serialized in Safetensors format. The 270M variant has a 0.3B parameter model size (Safetensors).

Microsoft notes that reproduced scores may differ slightly from reported benchmarks due to library version differences in PyTorch and Transformers.

Performance Claims

According to Microsoft, the Harrier models achieve state-of-the-art results on the Multilingual MTEB v2 benchmark as of the release date. The 27B model significantly outperforms the smaller variants: 74.3 vs. 69.0 and 66.5 respectively.

What This Means

Harrier fills a gap for production embedding models that handle long sequences (32K tokens) and multilingual content without reliance on proprietary APIs. The three-tier parameter design allows organizations to choose between efficiency (270M for edge deployment) and accuracy (27B for complex retrieval). The requirement for task-specific instructions during inference adds operational complexity but enables customization across different search and classification scenarios. Open-source availability means researchers can fine-tune variants for domain-specific embeddings without vendor lock-in.

Related Articles

model release

Microsoft launches seven models at Ignite; claims MAI Code 2 matches 96% of Claude Opus 5.5 at 5% of price

Microsoft unveiled seven AI models at Ignite on Nov. 17, aimed at coding, cybersecurity and other enterprise tasks. The company claims MAI Code 2 delivered 96% of Claude Opus 5.5's performance across four coding benchmarks at 5% of the price, using a promotional rate.

model release

Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license

Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.

model release

Reson8 releases 500M-parameter Rumble-2-Flash speech model as free Apache 2.0 open weights

Amsterdam-based Reson8 (Resonate Labs B.V.) has released Rumble-2-Flash, a 500 million-parameter transcription model covering more than 40 languages, as a free Apache 2.0 open-weights download. The company claims a 7.09% average word error rate across seven languages on Hugging Face's Open ASR Leaderboard multilingual tests.

model release

OpenAI releases GPT-Live-1.1 in the API, with standby mode billing idle WebRTC time at $0.02 per minute

OpenAI has released GPT-Live-1.1 in the API. The full-duplex voice model improves multilingual accents and transcription, and a new standby mode bills idle WebRTC time at $0.02 per minute instead of $0.05.

Comments

Loading...