Case Study Deskcasestudydesk.com
Developer Tools / CloudSourced

Case Study: Vercel's v0 saw a 40x latency improvement with Fireworks AI

Vercel Case StudySourced & dated by Case Study Desk
Key facts · TL;DR
Company
Vercel
Industry
Developer Tools / Cloud
Challenge
Slow, error-prone auto-fixing in the v0 generation pipeline
Headline result
A fine-tuned RL model pushed v0 error-free generation into the 90s

Key results

40x
Latency improvement
v0 auto-fixer vs gpt-4o-mini
93.87%
Error-free generation
v0-1.5-md model
8,130
Chars/sec
vercel-autofixer-01

The challenge

Vercel's v0 needed faster, more reliable code generation and auto-fixing than a general-purpose model provided, where multiple passes were sometimes required to fix a file.

The solution

Vercel worked with Fireworks AI to build a fine-tuned reinforcement learning model for v0's auto-fixer and composite generation.

Using a fine-tuned reinforcement learning model with Fireworks, we perform substantially better than SOTA. In our evaluation, Sonnet 3.5 compiled at 62%, and we got our error-free generation rate well into the 90s.

MU
Malte Ubl
CTO, Vercel

The results, in context

Vercel reported a 40x improvement in end-to-end latency for v0's auto-fixer model versus gpt-4o-mini, with the v0-1.5-md model reaching a 93.87% error-free generation rate and the vercel-autofixer-01 model running at 8,130 chars/sec. The company said its error-free generation rate moved well into the 90s versus 62% for Sonnet 3.5 in its evaluation.

Products used

Fireworks AI Fireworks AIFireworks AI Fine-tuningFireworks AI Reinforcement learning