Back to feed
Dev.to
Dev.to
6/18/2026
I Ran Five Small Multimodal Models on a Jetson. The Fastest One Was Not the Best Baseline.

I Ran Five Small Multimodal Models on a Jetson. The Fastest One Was Not the Best Baseline.

Short summary

Author benchmarked five multimodal models (Gemma, Qwen, SmolVLM, InternVL, Qwen-Omni) on NVIDIA Jetson for industrial edge AI. SmolVLM was fastest (12.8s) but lacked grounding for factory guidance. Gemma 4 E2B became the baseline—it balances speed (37.5s), structured action-card generation, and audit-trail compliance for maintenance, quality, and work-instruction workflows.

  • •Fastest model (SmolVLM2-2.2B, 12.8s latency) produced generic answers unsuitable for structured industrial guidance
  • •Gemma 4 E2B selected as baseline despite not being fastest—excels at audit trails, action boundaries, and workflow integration
  • •Model selection framework: local deployment + deterministic guards + audit compliance matters more than leaderboard speed

Generated with AI, which can make mistakes.

#ai-tools#ai-agents#research-breakthrough
Read full article at Dev.to

Is this a good recommendation for you?

Explore more

Model Showdown Round 7: Five Local Models vs. One Cloud Model on a Real Coding Task
Dev.to
Model Showdown Round 3: Ditching Ollama in Favor of llama.cpp
Dev.to
I shipped local LLM features two months ago. Production never ran them once.
Dev.to
Manufacturing intelligence with Amazon Nova Multimodal Embeddings
AWS Machine Learning Blog