Optimizing Vision-Language Models for Product Cataloging | Case study | Liquid AI

The challenge

A retail leader used large VLMs to catalog products—but models were slow, generic, and costly to fine-tune, creating deployment bottlenecks.

Key Obstacles

• Slow inference: Even quantized models lagged in production
• Poor specialization: Struggled with structured data extraction
• Complex deployment: Months of tuning needed for accuracy

Our solution

Liquid fine-tuned smaller, specialized VLMs for cataloging, using our Edge SDK to optimize both inference speed and accuracy.

The results

Faster, more accurate cataloging with 65% lower deployment time.

• 65% faster time-to-production
• Higher accuracy than larger generic models
• 50% lower compute/memory needs
• Seamless pipeline from fine-tuning to deployment