Eximius Labs

This is Eximius Labs' home for open model weights and multimodal embedding research. Everything here gives developers and researchers a production-ready starting point for cross-modal retrieval and semantic search, backed by the research behind the Fusion Embedding stack. The technical report is on arXiv: https://arxiv.org/abs/2607.18666.

The Fusion Embedding family

One embedding space for text, images, video, audio — and, with its first sense pack, thermal — built on a single design constraint: the base model is never retrained. Both generations are live on the public MTEB leaderboards for audio (MAEB) and video (MVEB), evaluated through the official mteb harness anyone can rerun.

fusion-embedding-1-2b-preview (v0.3) — generation 1 trains a 16.4M-parameter connector between a frozen Qwen2.5-Omni audio tower and a byte-frozen Qwen3-VL-Embedding-2B base: under 1% of the model. AudioCaps audio-to-text R@10 0.741, at the frontier of models that keep their base frozen. Aligned on audio-text only, it retrieves images from sound at 29x chance with zero paired audio-visual training data.

fusion-embedding-2-2b-preview (v0.2) — generation 2 adds modality-gated adapters inside the frozen decoder layers (+44.2M parameters), opened only for audio inputs. For text, images, and video the adapters never execute and outputs are bit-for-bit identical to the base model, verified after every training run. The family's best text-to-audio results on every benchmark: AudioCaps R@10 0.785, Clotho zero-shot 0.485.

fusion-embedding-2-ember (v0.1) — Ember is the thermal sense for fusion-embedding-2 and the first sense pack: a separately loadable adapter pack that adds a modality while every existing vector stays valid. Text-to-thermal retrieval R@10 0.785 on a 2,000-image gallery against a 0.224 frozen baseline. Sense packs are named for the physical trace their sensor reads; Ember reads heat.

How it works

A perceiver-resampler connector maps frozen audio-tower features into the base's input space; gated bottleneck adapters add in-layer capacity per modality. Gates are mutually exclusive by input modality, so non-gated modalities execute the base's exact computation. A single forward pass yields nested Matryoshka embeddings truncatable to {2048, 1536, 1024, 512, 256, 128, 64} dimensions. Training runs in hours on a single GPU; connectors and packs ship as small artifacts with the frozen towers downloaded separately.

import mteb
model = mteb.get_model("EximiusLabs/fusion-embedding-2-2b-preview")

Training data

The audio corpus is ~500K open audio-caption pairs (AudioCaps and AudioCaps 2.0, FSD50K, WavCaps/AudioSet-SL, a LAION-FreeSound subset); Ember trains on IR-TD real thermal imagery with restricted sources removed. Evaluation sets are excluded from training by clip id at ingestion, and every reported number traces to a result record in the open repository.

Status

English-language research previews. Model weights are released under CC-BY-NC-4.0 (Ember under its own research license; see each model card). Sound-event retrieval is the strongest axis; speech and music are undertrained in the current corpus. We release early to invite research and feedback on unified multimodal embedding spaces.

eximiuslabs.com · github.com/Eximius-Labs