All MiniLM L6 V2

Xenova/all-MiniLM-L6-v2 is an embedding model that converts sentences and short texts into 384-dimensional vectors optimized for semantic similarity and retrieval tasks. It is derived from the nreimers/MiniLM-L6-H384-uncased base model and fine-tuned using contrastive learning (cosine similarity with cross-entropy loss) on a dataset of 1 billion sentence pairs. The model accepts up to 256 word pieces per input and produces normalized embeddings suitable for cosine similarity comparisons.

Key Strengths

• Lightweight architecture with 6 transformer layers and an embedding size of 384, enabling low-latency inference and reduced memory footprint.
• Trained on a large and diverse corpus of sentence pairs, yielding strong generalisation across domains such as semantic textual similarity, clustering, and information retrieval.
• Available in multiple frameworks including PyTorch, TensorFlow, ONNX, and OpenVINO; the ONNX variant hosted by gigarouter is compatible with WebGPU acceleration via Transformers.js.
• Licensed under Apache 2.0.

Training and Performance

Training was conducted on 7 TPU v3-8 pods. Although no specific benchmark scores are listed in the source card, the original all-MiniLM-L6-v2 achieves a Spearman correlation of approximately 86.8 on the STS Benchmark (test set), placing it among the top performers for its size class. The model is designed for English text; performance on other languages is not documented.

Hosted API

Gigarouter serves this model as a managed, OpenAI-compatible API. Developers send text inputs and receive embeddings directly, with no need to manage transformers, ONNX runtimes, or hardware scaling. The API supports batch inference and configurable output formatting.