Model catalog

Every model candding ships, generated from the registry: the dense, sparse, multi-vector and rerank catalogs, each row with its own per-device status.

The tables are generated from the registry by just docs-models, from the JSON that candding list-models --json prints.

Model	Family	Dim	Pooling	Max length	Dtypes	CPU	Metal	CUDA	License
BAAI/bge-small-en-v1.5	BERT	384	cls	512	f32f16bf16	verified	verified	verified	MIT
BAAI/bge-base-en-v1.5	BERT	768	cls	512	f32f16bf16	verified	verified	verified	MIT
BAAI/bge-large-en-v1.5	BERT	1024	cls	512	f32f16bf16	verified	verified	verified	MIT
BAAI/bge-small-zh-v1.5	BERT	512	cls	512	f32f16bf16	verified	verified	verified	MIT
sentence-transformers/all-MiniLM-L6-v2	BERT	384	mean	256	f32f16bf16	verified	verified	verified	Apache-2.0
thenlper/gte-base	BERT	768	mean	512	f32f16bf16	verified	verified	verified	MIT
thenlper/gte-large	BERT	1024	mean	512	f32f16bf16	verified	verified	verified	MIT
mixedbread-ai/mxbai-embed-large-v1	BERT	1024	cls	512	f32f16bf16	verified	verified	verified	Apache-2.0
Snowflake/snowflake-arctic-embed-xs	BERT	384	cls	512	f32f16bf16	verified	verified	verified	Apache-2.0
Snowflake/snowflake-arctic-embed-s	BERT	384	cls	512	f32f16bf16	verified	verified	verified	Apache-2.0
Snowflake/snowflake-arctic-embed-m	BERT	768	cls	512	f32f16bf16	verified	verified	verified	Apache-2.0
Snowflake/snowflake-arctic-embed-l	BERT	1024	cls	512	f32f16bf16	verified	verified	verified	Apache-2.0
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	BERT	384	mean	128	f32f16bf16	verified	verified	verified	Apache-2.0
intfloat/multilingual-e5-small	BERT	384	mean	512	f32f16bf16	verified	verified	verified	MIT
Qwen/Qwen3-Embedding-0.6B	Qwen3	1024	last_token	8192	f32f16bf16	verified	verified	verified	Apache-2.0
Qwen/Qwen3-Embedding-4B	Qwen3	2560	last_token	8192	bf16f32f16	verified	verified	verified	Apache-2.0
Qwen/Qwen3-Embedding-8B	Qwen3	4096	last_token	8192	bf16f32f16	untested	verified	untested	Apache-2.0
google/embeddinggemma-300m	Gemma3	768	mean	2048	f32f16bf16	verified	verified	untested	gemma
intfloat/multilingual-e5-base	XLM-RoBERTa	768	mean	512	f32f16bf16	verified	verified	verified	MIT
intfloat/multilingual-e5-large	XLM-RoBERTa	1024	mean	512	f32f16bf16	verified	verified	verified	MIT
BAAI/bge-m3	XLM-RoBERTa	1024	cls	8192	f32f16bf16	verified	verified	verified	MIT
sentence-transformers/paraphrase-multilingual-mpnet-base-v2	XLM-RoBERTa	768	mean	128	f32f16bf16	verified	verified	verified	Apache-2.0
sentence-transformers/all-mpnet-base-v2	MPNet	768	mean	384	f32f16bf16	verified	verified	verified	Apache-2.0
jinaai/jina-embeddings-v2-small-en	JinaBERT	512	mean	8192	f32f16bf16	verified	verified	verified	Apache-2.0
jinaai/jina-embeddings-v2-base-en	JinaBERT	768	mean	8192	f32f16bf16	verified	verified	verified	Apache-2.0
jinaai/jina-embeddings-v2-base-de	JinaBERT	768	mean	8192	f32f16bf16	verified	verified	verified	Apache-2.0
jinaai/jina-embeddings-v2-base-zh	JinaBERT	768	mean	8192	f32f16bf16	verified	verified	verified	Apache-2.0
jinaai/jina-embeddings-v2-base-es	JinaBERT	768	mean	8192	f32f16bf16	verified	verified	verified	Apache-2.0
jinaai/jina-embeddings-v2-base-code	JinaBERT	768	mean	8192	f32f16bf16	verified	verified	verified	Apache-2.0
nomic-ai/nomic-embed-text-v1	NomicBERT	768	mean	8192	f32f16bf16	verified	verified	verified	Apache-2.0
nomic-ai/nomic-embed-text-v1.5	NomicBERT	768	mean	8192	f32f16bf16	verified	verified	verified	Apache-2.0
Snowflake/snowflake-arctic-embed-m-long	NomicBERT	768	cls	8192	f32f16bf16	verified	verified	verified	Apache-2.0
Alibaba-NLP/gte-modernbert-base	ModernBERT	768	cls	8192	f32f16bf16	verified	verified	verified	Apache-2.0
Alibaba-NLP/gte-multilingual-base	GTE (new)	768	cls	8192	f32f16bf16	verified	verified	verified	Apache-2.0
Alibaba-NLP/gte-base-en-v1.5	GTE (new)	768	cls	8192	f32f16bf16	verified	verified	verified	Apache-2.0

Dtypes: bold is the default the builder loads; plain text is verified, muted is untested, struck through is unsupported, each colored by the dtype's own Metal status so a processor-only refusal does not read as unsupported everywhere; hover a dtype for its own cpu/metal/cuda breakdown. F16 is refused on the CPU and verified on Metal for Qwen/Qwen3-Embedding-0.6B, Qwen/Qwen3-Embedding-4B, jinaai/jina-embeddings-v2-base-de and jinaai/jina-embeddings-v2-base-zh. BF16 is refused on the CPU for every row: candle has no CPU BF16 matmul.

Sparse models

Model	Head	Vocab	Max length	Phase	Dtypes	License
prithivida/Splade_PP_en_v1	splade_mlm	30522	512	3	f32f16bf16	Apache-2.0
BAAI/bge-m3	m3_linear	250002	8192	3	f32f16bf16	MIT
naver/splade-code-06B	splade_tied	151936	1024	3	f32f16bf16	NAVER Non-Commercial (Splade-Code)
jinaai/jina-embeddings-v2-small-en	mini_coil	19125	8192	3	f32f16bf16	Apache-2.0
sentence-transformers/all-MiniLM-L6-v2	bm42	30522	256	3	f32f16bf16	Apache-2.0
Qdrant/bm25	bm25	n/a	n/a	3	f32f16bf16	Apache-2.0

n/a marks an axis that does not exist for that row rather than one nobody has checked yet — Qdrant/bm25 has no encoder, no vocabulary and no device or numeric format anywhere in its own computation. Dtypes read the same way as in the dense table. F16 is refused on the CPU and verified on Metal for BAAI/bge-m3 and jinaai/jina-embeddings-v2-small-en. BF16 is refused on the CPU for every row except Qdrant/bm25, which has no device axis: candle has no CPU BF16 matmul.

Multi-vector models

Model	Head	Width	CLS row	Normalized	Score reduction	Query length	Max length	Phase	Dtypes	License
BAAI/bge-m3	m3_colbert	1024 → 1024	no	yes	mean_by_query_length	n/a	8192	4	f32f16bf16	MIT
colbert-ir/colbertv2.0	stanford_colbert	768 → 128	yes	yes	sum	32	180	4	f32f16bf16	MIT
answerdotai/answerai-colbert-small-v1	stanford_colbert	384 → 96	yes	yes	sum	32	300	4	f32f16bf16	Apache-2.0
jinaai/jina-colbert-v2	jina_colbert	1024 → 128	yes	yes	sum	32	300	4	f32f16bf16	CC-BY-NC-4.0

width is the projection's input and output row length; n/a in Query length marks a checkpoint whose own head fixes no such length at all, the same way it marks a missing axis in the sparse table above. Dtypes read the same way as in the dense table. BF16 is refused on the CPU for every row: candle has no CPU BF16 matmul.

Rerank models

Model	Family	Head	Score	Input	Max length	Phase	Dtypes	License
cross-encoder/ms-marco-MiniLM-L6-v2	BERT	bert_pooler	raw logit	pair, longest first	512	5	f32f16bf16	Apache-2.0
jinaai/jina-reranker-v1-tiny-en	JinaBERT	bert_pooler	probability	pair, longest first	512	5	f32f16bf16	Apache-2.0
BAAI/bge-reranker-base	XLM-RoBERTa	roberta_classification_head	raw logit	pair, query capped at 3/4	512	5	f32f16bf16	MIT
BAAI/bge-reranker-v2-m3	XLM-RoBERTa	roberta_classification_head	raw logit	pair, query capped at 3/4	512	5	f32f16bf16	Apache-2.0
Qwen/Qwen3-Reranker-0.6B	Qwen3	logit_score	raw logit	prompt, Qwen3 reranker template	8192	5	f32f16bf16	Apache-2.0
Qwen/Qwen3-Reranker-4B	Qwen3	logit_score	raw logit	prompt, Qwen3 reranker template	8192	5	f32f16bf16	Apache-2.0
Qwen/Qwen3-Reranker-8B	Qwen3	logit_score	raw logit	prompt, Qwen3 reranker template	8192	5	f32f16bf16	Apache-2.0
jinaai/jina-reranker-v2-base-multilingual	XLM-RoBERTa (flash), learned positions	roberta_classification_head	probability	windowed pair, best window	1024	5	f32f16bf16	CC-BY-NC-4.0

Tooltips give the values candding list-models prints, and the rerank page explains each score and input. Dtypes read the same way as in the dense table. F16 is refused on the CPU and verified on Metal for cross-encoder/ms-marco-MiniLM-L6-v2, jinaai/jina-reranker-v1-tiny-en, BAAI/bge-reranker-base, BAAI/bge-reranker-v2-m3, Qwen/Qwen3-Reranker-0.6B and jinaai/jina-reranker-v2-base-multilingual. BF16 is refused on the CPU for every row: candle has no CPU BF16 matmul.

Reading the columns

Column	Tables	Meaning
Pooling	dense	how per-token states become one vector, `cls`, `mean` or `last_token`; see pooling and templates
Max length	all	the default truncation, special and template tokens included; the builder caps any length at the model's position limit
Dtypes	all	F32, F16 and BF16, styled as the footnote under the dense table explains; hover a dtype for its CPU, Metal and CUDA statuses
CPU, Metal, CUDA	dense	the model's status on each device, as below
Phase	sparse, multi-vector, rerank	the roadmap phase that added the row
License	all	the licence of the weights on Hugging Face, which governs their use; candding's own is on the API page
Head	sparse, multi-vector, rerank	the head the row runs; the sparse, multi-vector and rerank pages explain each
Vocab	sparse	the row's index width; `n/a` for `Qdrant/bm25`, as are its max length and dtypes
Width	multi-vector	the projection's input and output width; every output row has the second
CLS row, Normalized	multi-vector	whether the classification position survives as a row, and whether every row is L2-normalized
Score reduction	multi-vector	how `late_interaction_score_for` reduces the summed score: `sum`, or `mean_by_query_length`
Query length	multi-vector	the fixed row count a query is padded to; `n/a` for a head that fixes none
Family	rerank	the family the checkpoint runs on; for XLM-RoBERTa (flash) it adds how the encoder numbers tokens
Score	rerank	what `RerankScore::value()` returns: `raw logit`, a difference of two logits for a `logit_score` head, or `probability`, after a sigmoid
Input	rerank	`pair` with its truncation, `longest first` or `query capped at 3/4`; `prompt` with its template; or `windowed pair, best window`; the rerank page explains each

Statuses

Status	Meaning
verified	the device's check passed in that dtype, as testing defines it
pending	a run is planned
untested	no run has reached the registry

Testing has the tolerance each dtype is held to.