skip to content
gigarouter gigarouter
models / specialist model · coming soon

EBind Full

encord-team/ebind-full

published Nov 2025 · updated Nov 2025

EBind Full is a multimodal embedding model that projects image, video, audio, text, and 3D point clouds into a shared embedding space for cross-modal similarity computations.

status
coming soon
API providers
0
downloads / mo
62
license
cc

specs

TaskMultimodal Embedding
ArchitectureEnsemble of Perception Encoder, ImageBind, and Uni3D
Parameters1.8 billion
Embedding Size1024
LicenseCC-BY-NC-SA 4.0

about this model

EBind is a multi-modal embedding model that projects image, video, audio, text, and 3D point cloud inputs into a shared 1024-dimensional embedding space for cross-modal similarity computation. It builds on three foundation models: the Perception Encoder, ImageBind, and Uni3D. Audio and 3D point cloud embeddings are projected via MLPs into the Perception Encoder's embedding space, producing unit-norm embeddings directly usable for cosine similarity comparisons.

The model achieves strong performance despite its compact 1.8B-parameter size, outperforming models 4 to 17 times larger across 13 benchmarks. This efficiency is enabled by a carefully curated dataset combining 6.7M fully-automated multimodal quintuples, 1M semi-automated human-annotated triples, and 3.4M pre-existing captioned items. Training requires only a single GPU in hours rather than days.

Diagram showing EBind architecture: Perception Encoder, ImageBind, and Uni3D encoders feeding into a shared embedding space via MLP projections for audio and 3D point clouds. Summary plot showing average benchmark performance across 13 evaluations versus model size, demonstrating EBind's competitive performance relative to models 4 to 17 times larger. Table 1: Retrieval benchmark results comparing EBind against larger models across multiple datasets. Table 2: Zero-shot classification benchmark results comparing EBind against larger models.

Key strengths include:

  • Supports five modalities (image, video, audio, text, 3D point clouds) in a single model.
  • Outperforms models 4 to 17 times larger on 13 benchmarks, including retrieval and zero-shot classification tasks.
  • Trainable on a single GPU in hours, not days.
  • Produces unit-norm embeddings for direct cosine similarity computation.

The model is published under the CC-BY-NC-SA 4.0 license. For further details, see the GitHub repository and the paper.

best for

FAQ

What modalities does EBind Full support?

It supports image, video, audio, text, and 3D point clouds.

What is the output embedding size?

The model produces 1024-dimensional unit-norm embeddings.

How many parameters does the model have?

EBind Full has approximately 1.8 billion parameters.

What is the license of the model?

It is released under the CC-BY-NC-SA 4.0 license.

How can I use this model via the gigarouter API?

Send requests to the OpenAI-compatible endpoint with your API key; see gigarouter documentation for details.

not yet live

We're benchmarking and onboarding EBind Full as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.

related specialist model models

compare all →