Supported models

rembed covers eight model architectures — six BERT-style encoders (BERT, DistilBERT, MPNet, RoBERTa including XLM-RoBERTa, ModernBERT, nomic-embed) and two decoder-derived embedders (Qwen3-Embedding and EmbeddingGemma) — across four tokenizer families (WordPiece, byte-level BPE, SentencePiece Unigram, Gemma byte-fallback BPE). Pass any of them to Load by Hugging Face id.

Every model below ships a committed golden and is validated in CI against an independent ONNX Runtime (or PyTorch) reference. "fp32 vs reference" is the worst-case max-absolute-difference; the int8 column is the worst-case cosine of the weight-only int8 path vs fp32.

Model	Pooling	fp32 vs reference	int8 (weight-only)
sentence-transformers/all-MiniLM-L6-v2	mean	1.5e-7	≥ 0.9991
sentence-transformers/all-MiniLM-L12-v2	mean	1.9e-7	in bounds
sentence-transformers/paraphrase-MiniLM-L3-v2	mean	< 1e-4	—
sentence-transformers/multi-qa-MiniLM-L6-cos-v1	mean	2.1e-7	≥ 0.998
BAAI/bge-small-en-v1.5	cls	< 1e-4	in bounds
BAAI/bge-base-en-v1.5	cls	7.4e-7	≥ 0.995
thenlper/gte-small	mean	2e-3¹	—
thenlper/gte-base	mean	3.8e-6	≥ 0.988
sentence-transformers/all-mpnet-base-v2	mean	3.3e-7	≥ 0.9978
sentence-transformers/paraphrase-mpnet-base-v2	mean	1.2e-6	≥ 0.9945
sentence-transformers/all-distilroberta-v1	mean	3.2e-7	≥ 0.9985
sentence-transformers/multi-qa-distilbert-cos-v1	mean	2.5e-7	≥ 0.999
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	mean	7e-7	≥ 0.9995
intfloat/multilingual-e5-small	mean	1.7e-7	≥ 0.9995
intfloat/multilingual-e5-base	mean	2.9e-7	≥ 0.999
Snowflake/snowflake-arctic-embed-s	cls	2.5e-7	≥ 0.995
nomic-ai/modernbert-embed-base	mean	< 1e-4	≥ 0.998
nomic-ai/nomic-embed-text-v1.5	mean	< 1e-4	≥ 0.996
Qwen/Qwen3-Embedding-0.6B	last-token	< 1e-4	≥ 0.997
google/embeddinggemma-300m	mean + Dense	< 1e-4	≥ 0.998

¹ gte-small ships as F16 weights: 2e-3 max-abs, but cosine ≥ 0.9999 and mean abs ≤ 2e-4.

Other checkpoints of these architectures are expected to work without a committed golden — the remaining e5 sizes, larger BGE/GTE, msmarco families, other BERT/DistilBERT sentence-transformers, and the larger Qwen3-Embedding (4B/8B, same architecture — run them cgo-free with WithDiskWeights on a machine that can't hold the fp32 weights in RAM).

Choosing a quantization mode

• fp32 — matches the reference within 1e-4. The default.
• Weight-only int8 (WithInt8) — cosine ≥ 0.988 across every model above; the safe quantized choice.
• Full int8 (WithInt8Activations) — fastest on AVX-VNNI CPUs, but activation error is per-checkpoint, not per-architecture (worst golden cosine ranges from ≥ 0.999 down to ~0.95 on some checkpoints, e.g. nomic-embed 0.953, bge-base 0.959). Not recommended for the RoBERTa family, ModernBERT, nomic-embed, or Qwen3 last-token pooling — prefer weight-only int8 when in doubt.

Retrieval prefixes

rembed embeds exactly the text you pass — it adds no prefixes. Models trained with instruction/role prefixes expect you to add them yourself:

Model family	Prefix
e5	"query: " / "passage: "
BGE	"Represent this sentence for searching relevant passages: " (queries)
Qwen3-Embedding	"Instruct: {task}\nQuery:{text}" on queries only; documents bare
arctic-embed	its own query prefix

Tokenizer fidelity

The WordPiece tokenizer reproduces Hugging Face's BasicTokenizer exactly — NFD accent stripping, CJK space-padding, BERT's punctuation predicate, and _clean_text parity: control and format characters (Cc/Cf/Co/Cs), including U+200C ZWNJ, are removed (an earlier build that kept ZWNJ failed the Persian golden).

The SentencePiece tokenizer (the XLM-RoBERTa family) has one deliberate divergence: on NFD-decomposed Hangul/kana it matches the SentencePiece C++ reference (recomposing NFD) rather than Hugging Face's fast tokenizer (which shreds Korean into jamo). A 65k-input fuzz run finds zero unintended mismatches.