SUVI: Single-Stream Vision AI for ARM

SUVI: Single-Stream Vision AI for Arm

SUVI, Scalable Unified Vector Intelligence, is a visually grounded multimodal architecture developed by Nakul Vyas at Heysuvi Labs, LLC.

Most vision-language systems use two large components: a language model and a separate vision tower. That works, but it adds memory use, runtime complexity, and deployment cost. SUVI takes a different route. It turns visual information into discrete tokens and processes image and language in one Mistral decoder stream.

The impact is simple: make grounded multimodal AI practical on Arm cloud and edge infrastructure without keeping a separate vision tower active for every request.

What we built

We deployed the Phase 3 SUVI architecture on Google Axion Arm infrastructure and optimized its inference path.

• A single-stream Mistral decoder combines visual tokens and text tokens.
• A learned visual bridge maps FSQ visual IDs into the decoder input stream.
• Images can be stored as compact 1 KB FSQ visual records.
• Cached visual records reconstruct the same visual prefix and caption token output as uncached processing.
• The architecture passed held-out Flickr30k causal-prefix grounding tests: real image prefixes outperformed shuffled, fixed, and random controls.

The public checkpoint is available on Hugging Face:

https://huggingface.co/nakulvyas21/suvi-phase3-flickr30k-arm

Arm optimization results

We tested SUVI on an 8-core Google Axion Arm VM and compared it with LLaVA v1.6 Mistral-7B, a dual-tower vision-language model.

Cached SUVI used 25.9% less peak memory and delivered 30.3% lower median request latency than the dual-tower comparator.

We also scaled cached SUVI across Arm CPU threads:

Why Arm

Arm is a strong fit for efficient AI deployment, but multimodal systems often bring large memory and runtime requirements. SUVI reduces the active inference path to compact visual IDs, a learned bridge, and one language decoder stream. This creates a cleaner route to multimodal AI on constrained cloud and edge systems.

The Arm work adds two practical optimizations:

1. FSQ visual caching: store reusable visual information as 1 KB records.
2. Eight-core Arm scaling: reduce request latency through efficient CPU parallelism.

What we learned

The major result is architectural. A single-stream design can retain image-specific visual grounding while reducing the resource burden associated with a separate always-on vision tower.

This project connects SUVI’s research architecture with a reproducible Arm deployment path, public code, public weights, and measured resource results.

Links

• Competition repository
• Public SUVI Phase 3 checkpoint
• IEEE CAI 2026 paper
• Heysuvi Labs, LLC

Built With

• diffusers
• gcp
• llava
• mistral
• peft
• python
• pytorch
• suvi
• transformers