Modal's free tier is $30/month of credit, and the useful translation is GPU-hours: at L4 pricing that's roughly 37 hours a month — more than an hour of GPU time a day, renewed monthly. For fine-tuning experiments or batch inference, that's genuinely a lot.
The way people waste it isn't compute, it's residency. Modal bills while a container is up, including warm-idle time you configured to dodge cold starts. A keep_warm=1 on a GPU function is ~720 hours of residency against a 37-hour budget — the credit evaporates in a day and a half with zero requests served.
The serverless discipline that keeps the tier free: let functions scale to zero, eat the cold start (seconds), and put the always-warm requirement on a CPU function if you must have it. Spend the credit on bursts, not on standing still.
What $30 covers per GPU type, with the math: https://toolfreebie.com/modal-serverless-gpu/
Top comments (0)