candding
Models

Model catalog

Every model candding ships, generated from the registry, with its pooling, length limit and backend verification status.

The table is generated from the registry by just docs-models; the same data drives the README table and the candding list-models --json command.

ModelFamilyDimPoolingMax lengthCPUMetalCUDALicense
BAAI/bge-small-en-v1.5BERT384cls512verifiedverifieduntestedMIT
BAAI/bge-base-en-v1.5BERT768cls512verifiedverifieduntestedMIT
BAAI/bge-large-en-v1.5BERT1024cls512verifiedverifieduntestedMIT
BAAI/bge-small-zh-v1.5BERT512cls512verifiedverifieduntestedMIT
sentence-transformers/all-MiniLM-L6-v2BERT384mean256verifiedverifieduntestedApache-2.0
thenlper/gte-baseBERT768mean512verifiedverifieduntestedMIT
thenlper/gte-largeBERT1024mean512verifiedverifieduntestedMIT
mixedbread-ai/mxbai-embed-large-v1BERT1024cls512verifiedverifieduntestedApache-2.0
Snowflake/snowflake-arctic-embed-xsBERT384cls512verifiedverifieduntestedApache-2.0
Snowflake/snowflake-arctic-embed-sBERT384cls512verifiedverifieduntestedApache-2.0
Snowflake/snowflake-arctic-embed-mBERT768cls512verifiedverifieduntestedApache-2.0
Snowflake/snowflake-arctic-embed-lBERT1024cls512verifiedverifieduntestedApache-2.0
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2BERT384mean128verifiedverifieduntestedApache-2.0

Reading the columns

  • Pooling is how per-token states become one vector: cls, mean or last_token; see pooling and templates.
  • Max length is the default truncation, taken from the repository's sentence-transformers configuration; the builder can raise it up to the model's position limit.
  • CPU, Metal, CUDA show verified when the golden suite (CPU) or the parity test (accelerators) passed, pending when a run is planned, and untested when nobody has run it yet; see testing.
  • License is the model's license on Hugging Face; candding itself is Apache-2.0 OR MIT.

How a model gets verified

A model is added to the registry with its pooling, templates and default length, a reference is generated with sentence-transformers in F32 on CPU, and the ten golden tests in candding/tests/golden_bert.rs must pass at cosine ≥ 0.9999 and max-abs ≤ 1e-4 before the entry lands. Metal and CUDA become verified when the parity test agrees with CPU at cosine ≥ 0.999 on every fixture input.

On this page