all-MiniLM
UKP Lab (TU Darmstadt)all-MiniLM maps a sentence or paragraph to a 384-dimensional dense vector for semantic search, clustering and retrieval-augmented generation. It runs a six-layer distilled MiniLM encoder contrastively fine-tuned on more than a billion sentence pairs, and truncates input beyond 256 word pieces. A paraphrase-multilingual sibling carries the same architecture across fifty-odd languages. The Sentence-Transformers project ships both.
Scored as the model family, not the sentence-transformers library, which is a separate product surface with its own PyPI channel. The two declared checkpoints are the ones the project fronts; ONNX, OpenVINO and quantized re-uploads published by third parties are not counted. Verified 2026-09-11 via the two Hugging Face model cards and the Hub API.
Openness
5 high confidence- weights
- open(Apache-2.0 ungated safetensors, PyTorch, ONNX and OpenVINO files)
- data
- open(all-MiniLM-L6-v2's 1.17B-pair corpus is a concatenation of named public datasets with per-dataset sampling weights in data_config.json)
- code
- open(train_script.py ships in the all-MiniLM-L6-v2 repository
- license
- Apache-2.0(OSI)
The headline checkpoint releases the whole contrastive run: the corpus is a list of named public datasets, the sampling weights ship as data_config.json, and the training script sits beside the weights under Apache-2.0. The MLM pretraining of the underlying MiniLM encoder is Microsoft's and is not part of this release, and the multilingual sibling documents its distillation without shipping the script.
- https://huggingface.co/api/models/sentence-transformers/all-MiniLM-L6-v2 recorded 2026-09-11
license: apache-2.0; gated: false; private: false; model.safetensors, pytorch_model.bin, tf_model.h5, ONNX and OpenVINO files, plus train_script.py and data_config.json in the repository listing.
- https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/raw/main/README.md recorded 2026-09-11
Lists the training corpus dataset by dataset with pair counts totalling over 1 billion, says the sampling configuration is in data_config.json, and states "The full training script is accessible in this current repository: train_script.py".
- https://huggingface.co/api/models/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 recorded 2026-09-11
license: apache-2.0; gated: false; private: false; safetensors, ONNX and OpenVINO weights present.
Adoption
5 high confidence254,324,263 downloads in the trailing 30 days for all-MiniLM-L6-v2 and 46,445,846 for the paraphrase-multilingual sibling. The Hugging Face universe sweep puts the 14 shipped MiniLM checkpoints in the sentence-transformers account at 310,843,911 combined. Third-party re-uploads and format conversions are excluded, and the figure is not the sentence-transformers PyPI package.
- https://huggingface.co/api/models/sentence-transformers/all-MiniLM-L6-v2 recorded 2026-09-11
downloads: 254324263.
- https://huggingface.co/api/models/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 recorded 2026-09-11
downloads: 46445846.
Capability
2 high confidenceNarrow and dated by the category rungs: single-language for the headline checkpoint, short-context, and a 2021 distillation generation. One rung below E5, which reaches the same short context with published multilingual retrieval numbers.
- https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/raw/main/README.md recorded 2026-09-11
States "By default, input text longer than 256 word pieces is truncated", describes English contrastive fine-tuning of nreimers/MiniLM-L6-H384-uncased, and publishes no retrieval benchmark table.
- https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2/raw/main/README.md recorded 2026-09-11
Declares 50+ language codes and a 384-dimensional sentence-similarity output, with no retrieval benchmark table.
Verified 2026-09-11