Multilingual search teams
Evaluate one model family across retrieval workloads that span many languages.
Independent tool overview
Harrier at a glance
Harrier is Microsoft's open-source family of multilingual text-embedding models for retrieval, RAG, semantic similarity, clustering, classification and reranking.
Visit the official Harrier site ↗
- Best for
- Multilingual semantic search and RAG retrieval
- Model sizes
- 270M, 0.6B and 27B parameters
- Maximum input
- 32,768 tokens
- Embedding dimensions
- 640, 1,024 or 5,376 by model
- License
- MIT
- Availability
- Model weights on Hugging Face; infrastructure not included
Overview
What Harrier is
Harrier turns text into dense numerical vectors that applications can compare for semantic search, retrieval-augmented generation, clustering and related tasks. It is a model family for developers, not a chatbot or a hosted end-user search product.
Microsoft released three sizes under the MIT license: 270M parameters for lighter deployments, 0.6B for a stronger quality-to-compute balance, and 27B for maximum benchmark performance. All support inputs up to 32,768 tokens, but each produces a different fixed embedding dimension.
At release, Microsoft reported that the 27B model ranked first on Multilingual MTEB v2 across 131 tasks. That makes Harrier an important candidate for multilingual retrieval evaluation, but benchmark leadership does not establish that it will be best on a specific company's documents, queries or latency budget.
Use cases
Who Harrier is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
RAG developers
Embed long documents and queries before vector search and generation.
Organizations self-hosting models
Run MIT-licensed embedding weights on controlled infrastructure rather than sending text to a proprietary API.
Teams balancing quality and compute
Benchmark three model sizes with different vector dimensions and resource requirements.
Capabilities
Core Harrier features
Three model sizes
Choose 270M, 0.6B or 27B parameters depending on the available hardware and retrieval target.
Multilingual embeddings
Represents text from many languages in a shared semantic vector space for cross-language and language-specific retrieval.
Long context
Accepts as many as 32,768 tokens per input according to Microsoft's model cards.
Dense normalized vectors
Uses last-token pooling and L2 normalization to produce fixed-size embeddings suitable for vector similarity search.
Retrieval and RAG
Supports query-to-document matching as the grounding layer before a generative model answers.
Broader embedding tasks
Can be evaluated for semantic similarity, clustering, classification, bitext mining and reranking.
Standard library support
Loads through Sentence Transformers or Transformers using the published Hugging Face repositories.
Permissive licensing
The MIT license permits broad commercial and internal use subject to its terms.
Process
How the Harrier workflow works
Step 1
Build a representative evaluation set
Collect real queries, relevant documents, hard negatives, languages and edge cases from the intended production workload.
Step 2
Select candidate sizes
Start with 270M or 0.6B for practical latency, then test whether the 27B model's quality gain justifies its compute and vector-storage cost.
Step 3
Embed queries and documents consistently
Follow the model card's current instructions and keep the same model, preprocessing and vector dimension across an index.
Step 4
Measure end-to-end retrieval
Compare recall, ranking quality, latency, memory, index size and downstream answer accuracy rather than relying only on public benchmarks.
Step 5
Version and monitor the index
Record the exact model revision and rebuild stored document vectors when changing the model, dimension or preprocessing.
Cost
Harrier pricing and free plan
Microsoft publishes the Harrier model weights under the MIT license, so there is no model license fee. Running the models still incurs compute, memory, storage and engineering costs. Harrier is not currently presented as a Microsoft-hosted embedding API with a public usage price.
Harrier OSS 270M
Free model weights
The smallest Harrier variant for constrained and lower-latency deployments.
- 270M parameters
- 640-dimensional vectors
- 32,768-token maximum
- Self-hosting costs apply
Harrier OSS 0.6B
Free model weights
A middle-size option intended to balance retrieval quality and deployment cost.
- 0.6B parameters
- 1,024-dimensional vectors
- 32,768-token maximum
- Self-hosting costs apply
Harrier OSS 27B
Free model weights
The flagship benchmark model with much higher hardware requirements.
- 27B parameters
- 5,376-dimensional vectors
- 32,768-token maximum
- Substantial inference and vector-storage costs
Pricing checked . Check current pricing at the source ↗
Assessment
Harrier strengths and limitations
Where it stands out
- Strong multilingual benchmark results at release
- Three sizes let teams trade quality for latency and infrastructure cost
- 32k-token input capacity supports long text chunks and documents
- MIT licensing and downloadable weights support private deployment
- Works with familiar Sentence Transformers and Transformers tooling
What to consider
- Harrier supplies model weights, not a complete vector database, RAG stack or managed production service
- Public benchmark scores do not guarantee the best retrieval quality on a specific domain or language mix
- The 27B model requires substantial accelerator memory and produces very large 5,376-dimensional vectors
- Switching between model sizes changes vector dimensions and requires re-embedding the document index
- Harrier is text-only; teams needing a shared text, image, audio and video space need a multimodal embedding model
- Microsoft's announcement says more than 100 languages while the Hugging Face model pages currently label 94, so teams should test every required language directly
Compare
Harrier alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Miscellaneous
Gemini Embedding 2
A managed multimodal embedding model for teams that need text, images, video and audio in one vector space.
Explore Gemini Embedding 2 →Coding
Together AI
A hosted AI platform for teams that prefer managed open-model inference instead of operating model infrastructure directly.
Explore Together AI →Coding
Ollama
A local model runner that may suit smaller private prototypes, though it is not a dedicated embedding benchmark family like Harrier.
Explore Ollama →Questions
Harrier FAQs
What is Microsoft Harrier?
Harrier is an open-source family of multilingual text-embedding models used to represent text as vectors for semantic search, RAG, similarity, clustering, classification and reranking.
Which Harrier model should I use?
The 270M model favors smaller deployments, 0.6B is the practical middle option, and 27B targets maximum benchmark quality. Test at least two sizes on real queries because the best choice depends on accuracy, latency, memory and index cost.
Is Harrier free?
The model weights are available under the MIT license with no model fee. You still pay for the hardware, hosting, vector database, storage and engineering needed to run a production retrieval system.
What is Harrier's context length?
Microsoft's model cards list a maximum of 32,768 tokens for all three Harrier OSS v1 variants.
What vector dimensions does Harrier use?
The 270M model outputs 640 dimensions, the 0.6B model outputs 1,024, and the 27B model outputs 5,376.
How many languages does Harrier support?
Microsoft's release post says more than 100 languages, while its Hugging Face pages currently carry a 94-language label. Validate performance for each language required by your application.
Can Harrier reduce RAG hallucinations?
Better retrieval can supply more relevant evidence to a generator and reduce unsupported answers, but the embedding model cannot guarantee factual output. Chunking, reranking, prompts, source quality and answer evaluation still matter.
Is Harrier a hosted API?
Microsoft currently publishes downloadable model weights and usage examples. The release announcement discusses future grounding services, but does not present Harrier OSS as a separately priced Microsoft-hosted API.
Bottom line
Our Harrier verdict
Harrier is a high-priority embedding candidate for teams building multilingual search or RAG with open weights. The 0.6B model is the sensible first benchmark for most teams; the 270M version favors constrained deployments, while 27B should be justified by measured retrieval gains. Treat the leaderboard result as a reason to test Harrier, not as a substitute for evaluating real queries.
Visit Harrier website ↗Similar Miscellaneous Tools
View all Miscellaneous AI tools →
Google Edge Eloquent
Google Edge Eloquent - Google AI Edge Eloquent — Free voice dictation app that turns messy speech into polished text, runs fully offline, no subscription, no usage caps
Lyra 2.0
Lyra 2.0 - NVIDIA's new model that turns text and camera paths into explorable 3D scenes
MAI-Transcribe-1 (Deprecated)
MAI-Transcribe-1 - Microsoft's speech-to-text model with best-in-class accuracy across 25 languages
Gemini 3.1 Flash TTS
Gemini 3.1 Flash TTS - Google's new speech model with inline tags for voice direction across 70+ languages
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.