One vector space for text, image, video, audio, and sensors, plus a perception layer and a memory layer. Open weights, self-hostable.
Eximius Labs builds open-weight multimodal models that run entirely on your own hardware. The stack has three complementary layers.
Fusion Embedding is the embedding layer. It maps text, images, video, audio, and sensor streams into a single shared vector space for retrieval, RAG, clustering, and cross-modal search. The weights are published and the models run on your own hardware.
Fusion Perception is the perception layer. Dense scene understanding and geometric place recognition on a frozen vision backbone, with a projector that drops its features into the Fusion Embedding space.
Engram is the memory layer. It indexes a robot's video, audio, and motion into that shared space on one clock and answers questions about it in plain language, including temporal reasoning that retrieval alone cannot do. Try it in the live playground.
We extend a state-of-the-art frozen base with new modalities and senses without modifying a single base weight. Text, image, and video vectors stay bit-for-bit identical to the base model, so gaining a modality never costs you a re-index.
| Model | Description |
|---|---|
| fusion-embedding-2 | Current line. Connector plus modality-gated deep adapters, 60.6M trained parameters. |
| fusion-embedding-1 | Connector-only architecture, 16.4M trained parameters. Final at v0.3. |
| Ember | Thermal sense pack: a separately loadable adapter pack that adds infrared to fusion-embedding-2 as a fifth modality, 44.2M trained parameters. |
| Tremor | Inertial sense pack: reads a robot's or wearable's accelerometer as language. Ships a general base and a Unitree-G1 head. |
Sense packs are named for the physical trace their sensor reads. Ember reads heat; Tremor reads motion. More senses are in training.
| Model | Description |
|---|---|
| fusion-perception-1 | Frozen DINO backbone plus light heads: geometric place recognition ("where am I"), segmentation, and a projector that makes those features language-searchable in the Fusion Embedding space. |
The open cross-modal memory layer for physical AI, built on Fusion Embedding. Index a robot's video, audio, and motion into one embedding space on a shared clock, then search and reason about it in plain language.
pip install engram-robomem
Code: Eximius-Labs/engram · PyPI · Playground
- Site: eximiuslabs.com
- Code: fusion-embedding · fusion-perception · engram — Apache-2.0
- Weights: huggingface.co/EximiusLabs
- Technical report: arXiv:2607.18666
