fastembed breaking changes
Every breaking API change we have recorded in fastembed on PyPI: 23 breaking changes in 1 version. Below are its newest version, removals and renames first, each with its source.
Newest change October 7, 2026

On this page
0.9.0
- Removed
fastembed.sparse.if_splade.iter_batch
fastembed.sparse.if_splade.iter_batch was removed in fastembed 0.9.0 pypi.org - Removed
fastembed.sparse.if_splade.load_tokenizer
fastembed.sparse.if_splade.load_tokenizer was removed in fastembed 0.9.0 pypi.org - Behavior changed
The model now loads model_v2.onnx and removes padding rows from its embeddings, so offline setups must download it first and images embedded in mixed-aspect-ratio batches should be re-embedded. github.com - Behavior changed
MUVERA encoding of an empty multivector now raises a ValueError instead of returning a result. github.com - Behavior changed
Truncation and padding side now come from the model's tokenizer config, which can change token outputs for affected models. github.com - Behavior changed
Passing a batch_size below 1 now raises a ValueError, so code that passed such values will fail. github.com - Behavior changed
Token limits were raised for some models on fresh downloads, so embeddings of long texts change. Cached models keep the old limit until deleted from the cache. github.com - Behavior changed
Embeddings from mean-pooling text models are now float32, so code relying on float64 output must adjust. github.com - Behavior changed
Built-in models no longer fall back to Google Cloud Storage when Hugging Face is unavailable. Cached GCS copies still load offline, and users should load each model once online before relying on offline use. github.com - Behavior changed
Center-crop preprocessing was changed, so image embeddings for some sizes differ from earlier versions. github.com
And 13 more in this version.
Check your code
These are the changes. To see which of them reach your code, scan a public repository, or have your coding agent call check_code with your manifest or lockfile over MCP. The API reads every change in every version, not only the newest. On Team and up, a fix pull request opens for the changes that reach a connected repository. Nothing merges or deploys.

Put your code on continuous management.
Scan a public repository now, or give your agent one line.
See how out of date your code is.
Tell your agent what changed since it was trained.
claude mcp add --transport http since https://api.since.dev/mcp