Skip to content

model2vec

Model Serving & MLOps

A technique that converts any sentence transformer into a small, fast static embedding model, reducing size up to 50× and speeding inference up to 500× with minimal performance loss

Track releases GitHub Website
Python Latest v0.8.2 · 1mo ago Security brief →

Features

  • Converts any Sentence Transformer into a compact static embedding model
  • Reduces model size by up to 50× and speeds inference up to 500×
  • Provides pre‑trained flagship models (e.g., potion-base-32M) for immediate use
  • Supports distillation of custom models from existing Sentence Transformers
  • Offers training extensions for fine‑tuning classification tasks on top of distilled models

Recent releases

View all 5 releases →
No immediate action
v0.8.2 Mixed

Training refactor + new features

Open
v0.8.1 Bug fix

Fixed POSIX path handling for Windows.

Full changelog

What's Changed

  • fix: Use posix path for Windows compatibility by @stephantul in https://github.com/MinishLab/model2vec/pull/314
  • increase version to 0.8.1 by @stephantul in https://github.com/MinishLab/model2vec/pull/315

Full Changelog: https://github.com/MinishLab/model2vec/compare/v0.8.0...v0.8.1

View release on GitHub
v0.8.0 Breaking risk
Breaking changes
  • Python 3.9 support deprecated
Notable features
  • Replace tokenizer with skeletoken
Full changelog

What's Changed

  • docs: Added CITATION.cff, updated citation info by @Pringled in https://github.com/MinishLab/model2vec/pull/293
  • docs: Added modelcard logo by @Pringled in https://github.com/MinishLab/model2vec/pull/294
  • docs: Update default model for example by @Pringled in https://github.com/MinishLab/model2vec/pull/299
  • chore: Deprecated Python 3.9 support by @Pringled in https://github.com/MinishLab/model2vec/pull/301
  • chore: Update CI to run on external PRs, fix test import by @Pringled in https://github.com/MinishLab/model2vec/pull/303
  • feat: replace tokenizer code with skeletoken by @stephantul in https://github.com/MinishLab/model2vec/pull/302
  • feat: refactor persistence, remove from_sentence_transformers by @stephantul in https://github.com/MinishLab/model2vec/pull/305
  • fix: refix tokenizer with added token shenanigans by @stephantul in https://github.com/MinishLab/model2vec/pull/304
  • fix: if layers == 0, layers were not initialized by @stephantul in https://github.com/MinishLab/model2vec/pull/306
  • feat: expose random seed, small typing fix by @stephantul in https://github.com/MinishLab/model2vec/pull/308
  • feat: add util for tokenizer pad id by @stephantul in https://github.com/MinishLab/model2vec/pull/310
  • docs: Update results to MTEB V2 by @Pringled in https://github.com/MinishLab/model2vec/pull/311
  • chore: Bump version by @Pringled in https://github.com/MinishLab/model2vec/pull/312

Full Changelog: https://github.com/MinishLab/model2vec/compare/v0.7.0...v0.8.0

View release on GitHub
v0.7.0 Breaking risk
Breaking changes
  • Rename PoolingType to PoolingMode
  • Deprecate apply_zipf and use_subword parameters
Notable features
  • Add vocabulary quantization
Full changelog

What's Changed

  • add support for passing weight to the loss functions by @volker48 in https://github.com/MinishLab/model2vec/pull/260
  • fix: padding token not recognized, update transformers by @stephantul in https://github.com/MinishLab/model2vec/pull/265
  • Fix tag train documentation by @Lhemamou in https://github.com/MinishLab/model2vec/pull/269
  • chore: Added python 3.13 to pyproject and CI by @Pringled in https://github.com/MinishLab/model2vec/pull/270
  • feat: add classifier freezing by @stephantul in https://github.com/MinishLab/model2vec/pull/274
  • fix: remove windows tests by @stephantul in https://github.com/MinishLab/model2vec/pull/277
  • feat: add configurable pad token by @stephantul in https://github.com/MinishLab/model2vec/pull/276
  • feat: faster loading if model already cached by @stephantul in https://github.com/MinishLab/model2vec/pull/278
  • feat: add vocabulary quantization by @stephantul in https://github.com/MinishLab/model2vec/pull/271
  • fix: load faster, make quantization better by @stephantul in https://github.com/MinishLab/model2vec/pull/279
  • fix: F rule, A rule, update ruff by @stephantul in https://github.com/MinishLab/model2vec/pull/281
  • feat: Added embedding_dtype and vocabulary_quantization to config by @Pringled in https://github.com/MinishLab/model2vec/pull/280
  • fix: Disable MPS for Torch versions >=2.8.0 by @Pringled in https://github.com/MinishLab/model2vec/pull/287
  • feat: Add configurable pooling for distillation by @Pringled in https://github.com/MinishLab/model2vec/pull/288
  • chore: Deprecate apply_zipf and use_subword parameters by @Pringled in https://github.com/MinishLab/model2vec/pull/289
  • chore: Rename PoolingType to PoolingMode by @Pringled in https://github.com/MinishLab/model2vec/pull/290
  • docs: Update main docs by @Pringled in https://github.com/MinishLab/model2vec/pull/291
  • chore: Bump version by @Pringled in https://github.com/MinishLab/model2vec/pull/292

New Contributors

  • @volker48 made their first contribution in https://github.com/MinishLab/model2vec/pull/260
  • @Lhemamou made their first contribution in https://github.com/MinishLab/model2vec/pull/269

Deprecation warnings ⚠️

  • apply_zipf and use_subword are now officially deprecated from distill

Full Changelog: https://github.com/MinishLab/model2vec/compare/v0.6.0...v0.7.0

View release on GitHub

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

About

Stars
2,150
Forks
121
Languages
Python Jupyter Notebook Makefile
View on GitHub Homepage Documentation

Install & Platforms

Install via
pip

Community & Support

Discord

Similar tools

Shdoc
CCC
Bullstudio v2

Beta — feedback welcome: [email protected]