NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #1556 most downloaded on PyPI
Fast, light, accurate library built for retrieval embedding generation
Latest stable
0.9.1
Last release 29 days ago
24 Oct 2026
Release timing varies
gaps range from 2 weeks to 6 months
Most releases are documented
notes for 29 of 39 stable releases
1 version withdrawn
withdrawn after publishing
3 years old
47 releases · first in 2023
Drag across the ruler to compare, or tap a release to open it.
fastembed-gpu 0.9.1 is released alongside, with the same changes.
minishlab potion models with numpy instead of ONNX, 1.9x to 2.6x faster on documents with unchanged outputs by @joeinmmap option for the potion models to memory-map the embedding matrix: it loads instantly and lets parallel workers share one copy by @joeinValueError when a custom model with PoolingType.DISABLED returns per-token output instead of sentence embeddings by @joeinfastembed-gpu 0.9.1 is released alongside, with the same changes.Built-in models now download from Hugging Face only, without the Google Cloud Storage (GCS) fallback some of them had. The GCS buckets stay open, so o
token_count() loads only the tokenizer, not the ONNX model, when lazy_load=True by @Mirochill @joeinlightonai/LateOn and the multilingual lightonai/mLateOn by @tekumara @joeinstemmer and stopwords parameters to Qdrant/bm25 for languages without a Snowball stemmer, such as Polish by @Mohith26 @joeinibm-granite/granite-embedding-small-english-r2 by @Aryan-SINGH-GIT @joeinBAAI/bge-reranker-v2-m3 and its int8 version BAAI/bge-reranker-v2-m3-int8 by @joeink_sim values by @Pothan0 @joeinQdrant/stella-en-400M-v5-doc-onnx for documents and the small query encoders Qdrant/constella-zero and Qdrant/constella-nano that search its embeddings by @joein @Dylancouzonoutput_name to TextEmbedding.add_custom_model() to choose which ONNX output a custom model uses by @tuanzirwar @joeinlazy_load=True and parallel by @Mirochill @joeinparallel run is stopped early, for example with break by @FU-max-boop @joeinValueError for a batch_size below 1 by @adityaanikam @joeinQdrant/colpali-v1.3-fp16 (gemma) and the Unicom models (mit) by @joeinjinaai/jina-clip-v1 preprocessing, which was off by 1 px for some image sizes by @joeinValueError when MUVERA encodes an empty multivector by @Pothan0 @joeinfloat32 instead of float64 embeddings from mean-pooling text models, such as sentence-transformers/all-MiniLM-L6-v2 (regression in 0.6.1) by @joein.onnx_data weights from a huggingface_hub>=1.32 cache with onnxruntime>=1.24 by @joeinQdrant/colmodernvbert image embeddings in batches that mix aspect ratios, for portrait images, and for very thin images, and speed up its text queries (see Upgrade Notes) by @hulkbig @JesusMan0529 @joeinint64 indices from SparseEmbedding.from_dict({}) by @RAMZI0TO99 @joeinQdrant/bm25 query embeddings by @RAMZI0TO99 @joeincuda=False in parallel workers, which ran on the GPU when onnxruntime-gpu was installed (regression in 0.8.0) by @RAMZI0TO99 @joeindo_resize=False in ConvNeXT image preprocessing by @RAMZI0TO99 @joeinlog1p by @RAMZI0TO99 @joeinurl and hf sources. GCS copies cached by older versions still load offline or when Hugging Face is unreachable, with a FutureWarning: load the model once online, then delete its fast-<model name> dir.sentence-transformers/all-MiniLM-L6-v2 and Qdrant/bm42-all-minilm-l6-v2-attentions (now 256 tokens), and thenlper/gte-large and prithivida/Splade_PP_en_v1 (now 512). This applies to any fastembed version on a fresh download, and embeddings of texts over 128 tokens change. A cached model keeps the old limit until you delete it from the cache.Qdrant/colmodernvbert now loads model_v2.onnx and drops padding rows from its embeddings, so offline setups need to download it first. Re-embed images that were embedded in batches with images of other aspect ratios. The model needs onnxruntime>=1.19.LateInteractionMultimodalEmbedding.embed_image now defaults to batch_size=2 instead of 16, which also applies to Qdrant/colpali-v1.3-fp16.jinaai/jina-clip-v1 image embeddings change for some image sizes.Thanks to everyone who contributed to this release @adityaanikam @Aryan-SINGH-GIT @Dmheath1 @Dylancouzon @FU-max-boop @HuaTNA @hulkbig @JesusMan0529 @Mirochill @Mohith26 @Pothan0 @RAMZI0TO99 @tekumara @tuanzirwar @joein
Your coding agent can read these notes before it upgrades. Set up the MCP server →
BAAI/bge-small-en-v1.5 (the default model) and BAAI/bge-base-en-v1.5 now resolve to Qdrant/bge-small-en-v1.5-onnx-Q and Qdrant/bge-base-en-v1.5-onnx-Q
google/embeddinggemma-300m by @joeinnomic-ai/nomic-embed-vision-v1.5 and nomic-ai/nomic-embed-vision-v1.5-Q, which share an embedding space with nomic-ai/nomic-embed-text-v1.5 for text-to-image search by @Dylancouzonopensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte, which runs the model on documents only and encodes queries without inference by @joeinQwen/Qwen3-Embedding-0.6B and Qwen/Qwen3-Embedding-0.6B-Q (the quantized model requires onnxruntime>=1.23), plus PoolingType.LAST_TOKEN for custom models by @joeingoogle/siglip2-base-patch16-224 to TextEmbedding and ImageEmbedding by @joeinminishlab/potion-base-8M, minishlab/potion-retrieval-32M, and minishlab/potion-multilingual-128M by @stephantul @DylancouzonBAAI/bge-base-en-v1.5 and BAAI/bge-small-en-v1.5 to avoid a redirect that broke downloads behind some proxies (see Upgrade Notes) by @Harnas @rastagan-git @joeinjinaai/jina-embeddings-v2-base-de instead of fp16, which fails on onnxruntime>=1.23 (the download grows from 0.32 GB to 0.64 GB) by @joeinKeyError when loading a custom text model with different casing than it was registered with by @CODING-DARSH @joein(N, C, H, W) image arrays, which ran along the wrong axis by @serhiizghama @joeinconfig.json and special_tokens_map.json optional when loading a tokenizer by @libaojiang @joeinResize transform swapping height and width for non-square sizes by @Ramnath0521 @joeinadd_custom_model) failing when parallel is set by @joein @S0rryHorizonValueError on mixed-length batches for models whose tokenizer ships a fixed padding length, such as thenlper/gte-base (regression in 0.8.0) by @joein @mohmedmmBAAI/bge-small-en-v1.5 (the default model) and BAAI/bge-base-en-v1.5 now resolve to Qdrant/bge-small-en-v1.5-onnx-Q and Qdrant/bge-base-en-v1.5-onnx-Q. The cache dir name follows the repo ID's casing, so on case-sensitive filesystems (typically Linux) the existing cache isn't reused and both models download again once. Offline setups (HF_HUB_OFFLINE=1 or local_files_only=True) need to refresh their cache before upgrading. You can delete the old models--qdrant--bge-*-onnx-q dirs afterwards.jinaai/jina-embeddings-v2-base-de now loads onnx/model.onnx instead of onnx/model_fp16.onnx, so offline setups need to download it first.Thanks to everyone who contributed to this release @CODING-DARSH @Dylancouzon @Harnas @he-yufeng @libaojiang @mohmedmm @Ramnath0521 @rastagan-git @S0rryHorizon @serhiizghama @stephantul @joein
Thanks to everyone who contributed to this release @joein @amasolov @kacperlukawski
cuda=TrueThanks to everyone who contributed to this release @joein @amasolov @kacperlukawski
Thanks to everyone who contributed to this release @dancixx @joein and @tbung for all the reviews
enable_cpu_mem_arena onnx session option to handle onnxruntime memory allocation by @joeintoken_count method to text models by @joein @dancixxThanks to everyone who contributed to this release @dancixx @joein and @tbung for all the reviews
Nothing published for this version
Thanks to everyone who contributed to this release @kacperlukawski @generall @joein
Thanks to everyone who contributed to this release @kacperlukawski @generall @joein
# Changelog ## Features 🏎️ * #532 - improved warnings for the recently changed models by @joein * #522 - raise exceptions in case of incorrect pooling
embedding_size and cls method get_embedding_size for easier access to the model dim by @joeinspecific_model_path and local_files_only to embed with parallel > 1 by @joein# Changelog ## Features 🏎️ * #513 - New sparse embeddings model with semantic understanding: MiniCOIL (Qdrant/minicoil-v1) by @generall ## Fixes 🕸️ *
Qdrant/minicoil-v1) by @generall# Changelog ## Features 🏎️ * #490 - deprecate old archive struct when loading from custom urls in favour of model_name.tar.gz to ease adding custom mo
model_name.tar.gz to ease adding custom models by @joeinThanks everyone who contributed to the current release!
specific_model_path bypassing hf file structure by @I8dNLoThanks everyone who contributed to the current release!
# Change Log ## Fixes 🐛 * #439 - fix python3.13 installation by making onnx an optional dependency, keep onnxruntime mandatory by @joein
onnx an optional dependency, keep onnxruntime mandatory by @joein## Features 📖 #403 - Drop Python 3.8 support by @joein #404 - Add Python 3.13 support by @joein #406 - Improve models cache progress bar by @hh-space-
#403 - Drop Python 3.8 support by @joein #404 - Add Python 3.13 support by @joein #406 - Improve models cache progress bar by @hh-space-invader #422 - Add multi-GPU example by @hh-space-invader #425 - Provide user warning when specifying providers and CUDA by @hh-space-invader
#405 - Support jina embeddings v2 models by @hh-space-invader #408 - Add jina clip v1 model by @hh-space-invader #415 - Added support for the thenlper/gte-base model by @hh-space-invader #419 - Introduced parallel processing and pair-wise API for cross-encoders by @I8dNLo #429 - All models now support Hugging Face (hf) compatibility by @I8dNLo
#413 - Fix ColBERT model shape mismatch by @hh-space-invader
# Change Log ## Features 📖 * #380 - add NOTICE file to support models with restrictive licenses by @hh-space-invader @joein * #362 - unlock numpy v2 b
# Change Log ## Features 📢 * #366 - replace pystemmer with py-rust-stemmers by @I8dNLo
Thanks to everyone who contributed to the current release @celinehoang177 @I8dNLo @generall @hh-space-invader @n0x29a @joein
Thanks to everyone who contributed to the current release @celinehoang177 @I8dNLo @generall @hh-space-invader @n0x29a @joein
Nothing published for this version
Thanks to everyone who contributed to the current release @I8dNLo @Anush008 @mrscoopers
model_max_length down in the case of context windows larger than 512 by @I8dNLoThanks to everyone who contributed to the current release @I8dNLo @Anush008 @mrscoopers
https://github.com/qdrant/fastembed/pull/291 - Add Unicom models by @I8dNLo
Nothing published for this version
Nothing published for this version
Add support for jinaai/jina-embeddings-v2-base-de by @deichrenner in https://github.com/qdrant/fastembed/pull/270
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.3.0...v0.3.1
# Changelog ## Features 🪄 - #219 support for ImageEmbedding by @joein - #219 CLIP image and text embeddings by @joein - #246 Resnet50 by @I8dNLo - #24
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.2.7...v0.3.0
Various documentation, workflow and notebooks improvements by @NirantK @generall @arunppsg
Various documentation, workflow and notebooks improvements by @NirantK @generall @arunppsg
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.2.6...v0.2.7
feat: support mixedbread-ai/mxbai-embed-large-v1 by @yuvraj-wale in https://github.com/qdrant/fastembed/pull/158
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.2.5...v0.2.6
Make debugging easier: Add import statement for version debugging by @NirantK in https://github.com/qdrant/fastembed/pull/151
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.2.4...v0.2.5
Fix splade for single input by @generall in https://github.com/qdrant/fastembed/pull/148
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.2.3...v0.2.4
Check for existing files in cache dir before instantiating a model by @nleroy917 in https://github.com/qdrant/fastembed/pull/128
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.2.2...v0.2.3
docs: Describe how to change the model and how to just create embeddings by @KShivendu in https://github.com/qdrant/fastembed/pull/112
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.2.1...v0.2.2
Simplify imports: #110 by @Okabe-Rintarou-0 in https://github.com/qdrant/fastembed/pull/113
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.2.0...v0.2.1
use "with" to open JSON files by @dpjanes in https://github.com/qdrant/fastembed/pull/96
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.1.3...v0.2.0
Nothing published for this version
Add huggingface-hub dependency by @NirantK in https://github.com/qdrant/fastembed/pull/66
Full Changelog: https://github.com/qdrant/fastembed/compare/v0.1.1...v0.1.2
Nothing published for this version
New Models: Add support for v1.5 models by @NirantK in https://github.com/qdrant/fastembed/pull/25
Full Changelog: https://github.com/qdrant/fastembed/compare/0.0.5...0.1.0
Remove bulky transformers and torch dependencies, directly uses tokenizer fix errors by @generall in https://github.com/qdrant/fastembed/pull/13
transformers and torch dependencies, directly uses tokenizer fix errors by @generall in https://github.com/qdrant/fastembed/pull/13list_supported_models. Existing models supported:
DefaultEmbeddingFull Changelog: https://github.com/qdrant/fastembed/compare/0.0.3a1...0.0.5
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
One column per quarter.