Card: note the integration files are on the kw-preview branch too

Model card: add Sentence Transformers usage section

Add Sentence Transformers integration (custom module)

Evaluation integrity: report Common Voice under the training_datasets disclosure; remove internal notes

Add the evaluation-integrity section (Common Voice overlap disclosure)

Add self-hosted serving section (vLLM and SGLang plugins)

Cite the corpus paper (arXiv:2608.01560) at the emotion-regression note it explains

Add Tactus (touch sense pack) to the model family listing

Install fusion-embedding from PyPI

Add Engram playground link

Add Engram ecosystem section

Card: clean rebuild - single-row badges + RunPod on next row, fixed frontmatter

Card: badges on one row, Deploy on RunPod on next row (centered)

Card: fix line endings (clean LF; restore YAML frontmatter parsing)

Card: revert to dynamic RunPod badge

Card: static Deploy on RunPod badge (always renders)

Card: restore banner, add RunPod deploy

Card: add RunPod deploy section

Card: add RunPod Hub deploy badge + endpoint call instructions

Add v0.3 banner to model card

v0.3-preview: retrieval-tuned flagship

Card: separate model citation (Tonmoy) from paper citation (all authors)

Cite the arXiv paper (2607.18666)

card: link the Ember thermal sense pack

Card: MAEB nine-task section with the v0.2 rescore

config.json: restore the remote-code loading schema

v0.2: fine-tune on the expanded AudioCaps 2.0 pool (86,394 pairs, 1,600 steps)

inference.py: native video preprocessing (no qwen-vl-utils)

embed_video: native preprocessing (frame tensors and torchcodec path decode); drops the qwen-vl-utils dependency; bitwise-equal to the previous implementation and the raw base on identical inputs

inference.py: embed_video (base's own video path, gate-guarded)

Add embed_video: the frozen base's own video path (qwen-vl-utils preprocessing, official defaults), gate-guarded like images; bitwise-verified against the base and the reference loader

Card: AudioCaps T2A R@1 predecessor cell now recorded (0.280)

Card: positioning alt text matches the trained-parameters axis

Positioning figure: trained-parameters axis (matches the technical report)

inference.py: install URL points at the renamed family repo

Card: code links to the renamed family repo (Eximius-Labs/fusion-embedding); Clotho T2A R@1 predecessor cell now recorded (0.136)

AutoModel + trust_remote_code support (remote-code loading; existing loaders unchanged)

Card: precise cell count (8 of 12 release-protocol cells; every recorded text-to-audio direction)

inference.py: gate guard on embed_image (non-audio inputs must run with the gate closed)

model card: positioning chart visible in Evaluation

assets: positioning chart — distinct hues for the two generations

model card: positioning chart in cross-modal section

assets: unified-model positioning chart

assets: overview figure from the unified family generator

assets: overview figure in Eximius Labs palette

assets: overview figure — EOS chip inside pooling box

assets: overview figure — horizontal pooling arrow

assets: model overview figure v2 (layer-stack style)

model card: add model overview figure

assets: model overview figure