SUVI Phase 3 Flickr30k Arm Checkpoint
This is the public inference checkpoint for the SUVI Phase 3 visual bridge: a visually grounded, single-stream multimodal architecture developed by Nakul Vyas at Heysuvi Labs, LLC.
It contains the LoRA adapter, learned visual bridge, tokenizer files, and the frozen Phase 3 contract used by the Arm optimization repository.
Run it
Use the implementation and setup instructions in the
SUVI Arm Optimization repository.
The checkpoint uses mistralai/Mistral-7B-v0.1 as its base decoder and
stabilityai/sd-vae-ft-mse for image encoding.
Training and evaluation
The visual bridge was trained on Flickr30k captions. The Phase 3 architecture passed held-out Flickr30k causal-prefix tests: real image prefixes outperformed shuffled, fixed, and random controls.
This checkpoint is released to reproduce the SUVI visual-grounding and Arm deployment work. Its purpose is to establish that the single-stream visual path carries image-specific information. It is not positioned as a finished caption-quality model; more training is the next step for richer captions.
Flickr30k images and captions are not included in this repository.
Citation
@article{vyas2026suvi,
title={SUVI: Scalable Unified Vector Intelligence for Efficient Edge Deployment},
author={Vyas, Nakul},
journal={IEEE Conference on Artificial Intelligence},
year={2026},
url={https://ieeexplore.ieee.org/document/11536337}
}
For partnership or product inquiries, contact nvyas@heysuvi.com.
Model tree for nakulvyas21/suvi-phase3-flickr30k-arm
Base model
mistralai/Mistral-7B-v0.1