zen-embedding-2/README.md

license, base_model, library_name, pipeline_tag, language, tags

license	base_model	library_name	pipeline_tag	language	tags
apache-2.0	google/embeddinggemma-2	sentence-transformers	feature-extraction	multilingual	zen zen-embedding zenlm hanzo embedding sentence-transformers multimodal-embedding vision audio video retrieval matryoshka

Zen Embedding 2

One embedding space for text (including code), images, video and audio: any mix of them in, a 768-dimensional vector out, so a text query retrieves a screenshot, a clip or a code span directly. 8,192-token context (minutes of audio or video), Matryoshka truncation to 512, 256 or 128 dimensions, and encoders that load per modality: 270M parameters for text alone, 740M with vision and audio.

Repackaged from google/embeddinggemma-2 (Apache-2.0, Google DeepMind) at revision 914f7f89142e33e77833254d9c9b90c3cef7303b, weights unchanged. Not trained from scratch. The upstream card has the benchmarks, training data and safety notes.

Usage

import torch
from sentence_transformers import SentenceTransformer

dtype = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else torch.float32
model = SentenceTransformer("zenlm/zen-embedding-2", model_kwargs={"torch_dtype": dtype})

query = model.encode("task: code retrieval | query: where is the cache key built?")
doc = model.encode("title: cache.py | text: def make_key(args, kwargs): ...")
print(model.similarity(query, doc))

Run it in bfloat16 or float32, never float16: its activations overflow float16 into NaN.

Task prefixes steer the vector. Queries take task: search result | query: …, task: question answering | query: …, task: fact checking | query: … or task: code retrieval | query: …; documents take title: {title or none} | text: …; symmetric tasks put task: classification | query: …, task: clustering | query: … or task: sentence similarity | query: … on every input.

Mixed input marks each item's place with <|image|>, <|video|> or <|audio|>:

emb = model.encode({
    "text": "Checkout page after submit: <|image|> Screen recording of the failure: <|video|>",
    "image": ["checkout.png"],
    "video": "failure.mp4",
})

Text only, without the vision and audio encoders:

model = SentenceTransformer("zenlm/zen-embedding-2",
                            config_kwargs={"vision_config": None, "audio_config": None})

Truncate with model.encode(x, truncate_dim=256, normalize_embeddings=True); queries and documents must share a dimension.

Served on the Hanzo gateway (api.hanzo.ai) as zen-embedding-2.

License

apache-2.0. Upstream: google/embeddinggemma-2 (EmbeddingGemma 2) by Google DeepMind. This repository redistributes the weights unchanged under the same license.