harrier-oss-v1-0.6b

Our Take

A strong 0.596B-parameter dense embedding model from Microsoft. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing.

Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.

Harrier-oss-v1-0.6b uses a dense decoder-only architecture with last-token pooling and L2 normalization to produce 1,024-dimensional embeddings. Unlike mixture-of-experts (MoE) models, all 0.596B parameters are active during inference. This means a single forward pass consumes the full compute, but the model is small enough to fit comfortably on most consumer GPUs and even run on CPU with acceptable throughput.

Key specs:

• Parameters: 0.596B (596M)
• Embedding dimension: 1,024
• Context length: 32,768 tokens (confirmed from web sources)
• Pooling: Last-token pooling (left-padding handled automatically)
• Training: Contrastive learning + knowledge distillation from larger embedding models
• Framework compatibility: sentence-transformers, transformers

The context window of 32K tokens is unusually large for a model of this size. Most sub-1B embedders cap out at 512 or 8,192. This enables processing long documents, code files, or multi-turn conversation histories without chunking – a significant advantage for RAG pipelines that index long-form content.

The model is a text-only embedder suited for any task that requires converting text into a fixed-length vector. The official model card lists retrieval, clustering, semantic similarity, classification, bitext mining, and reranking. Here are concrete use cases relevant to local deployments:

• Multilingual semantic search: Given support for 94+ languages (including low-resource languages like Amharic, Yoruba, and Kurdish), you can build a single search index that handles queries and documents in different languages. Microsoft trained on a large mixture of multilingual datasets – this is not a tokenizer-level afterthought.

• RAG with long documents: Use the 32K context window to embed entire research papers, legal contracts, or medical records without sharding. Combine with a vector database like Chroma or LanceDB for local retrieval.

• Bitext mining and translation memory: Compare embeddings across parallel corpora to extract aligned sentence pairs – useful for custom machine translation filtering.

• Classification and clustering: Use the embeddings as features for logistic regression or k-means. The 1,024-dimensional vectors retain enough signal for fine-grained topic labeling.

• Reranking: Generate query-document scores (via dot product) to reorder initial retrieval results from a cheaper lexical index.

The model is not designed for generative tasks (no chat completion, no code generation). It outputs only embeddings.

This is where the model shines for practitioners on limited hardware.

At full precision (FP32), the model weights occupy approximately 2.4 GB (0.596B × 4 bytes). With typical overhead for tokenizers and intermediate activations, expect peak usage around 3–4 GB for a single batch.

• Q8_0 / Q8_0_M: ~1.5 GB – runs on any GPU with 4 GB VRAM (e.g., GTX 1060, RTX 3050).
• Q4_K_M: ~0.8 GB – the recommended quantization for most users. Runs on CPU with <1 GB RAM overhead, or on 2 GB GPUs (e.g., Jetson Nano, older laptop GPUs).
• Q2_K / Q3_K_M: ~0.5–0.6 GB – suitable for edge devices with extreme memory constraints; expect some accuracy degradation on MTEB score (estimated 1–2 points drop).

• GPU: Realistically runs on any NVIDIA GPU with at least 4 GB VRAM (RTX 3060, RTX 4060, RTX 4090). For 1,024-dimensional embedding of a 1K-token document, expect 200–500 tokens/second on a mid-range GPU like the RTX 3060, and 800+ tokens/second on an RTX 4090.
• CPU: A modern 8-core CPU (e.g., AMD Ryzen 7, Intel Core i7) can process ~50–100 tokens/second using ONNX or llama.cpp-based quantized variants. The 0.6B size makes it one of the few large-context embedders that still feel interactive on CPU.
• Apple Silicon: M1/M2/M3/M4 Macs (8 GB or more) can run the Q4 quantized model via llama.cpp or the mlx framework. Expect 100–200 tokens/second on M4 Max.

The quickest path is via Ollama. While the Harrier family is not yet in the official Ollama library, you can import the model from HuggingFace:

Or use Sentence Transformers directly:

For maximum performance on CPU, convert to GGUF format (e.g., Q4_K_M) and run via llama.cpp:

Then load with llama-embedding or the Python binding.

• Harrier has 5× the parameters, 8× the context window (32K vs 512), and a higher MTEB multilingual score (69.0 vs ~62).
• E5-small is smaller and faster on CPU, but limited to 512 tokens – requires chunking for long documents.
• When to choose Harrier: You need multilingual retrieval on documents longer than a paragraph, or want to avoid complex chunking strategies.

• Both are dense embedders with similar parameter counts and multilingual support.
• BGE-M3 supports 8192 context and 1024-dim embeddings, but its MTEB benchmark performance is slightly lower (~67.5 on similar tasks).
• Harrier has a 4× larger context window (32K vs 8K) and uses knowledge distillation – expect better accuracy on dense retrieval tasks, especially for long texts.
• When to choose Harrier: You need the extra 24K tokens of context for indexing lengthy documents, or you want the latest distilled checkpoint. BGE-M3 is a solid fallback if you have existing infrastructure.

For purely CPU deployments, the 0.6B parameter count of Harrier makes it a practical choice – most 7B-class embedders are unusable on CPU for real-time retrieval.