Pick Modal if
python teams with spiky GPU/CPU jobs who will read a per-second table.
Read the Modal pageModal vs vLLM
Our take
Pick Modal if python teams with spiky GPU/CPU jobs who will read a per-second table. Pick vLLM if platform and inference engineers who already have a GPU (or a kube cluster) and need an OpenAI-shaped endpoint. Skip Modal if people who want a hosted chat playground (Together AI) or a one-click image studio (Replicate). Skip vLLM if someone who wants a local ChatGPT window (Ollama, Jan) and will not run vllm serve.
vLLM
high-throughput LLM serving engine
Best for
Modal Run a containerized function on burst GPUs
vLLM Serving tokens at production QPS
Free plan
Modal $0 · / month + compute. Up to 3 seats. Included compute $30 / month. 1-day logs. 5 deployed crons
vLLM $0 · apache-2.0. You pay the GPU, electricity, and (if you use them) model licenses
Lowest paid
Modal CPU $0.0000131/mo
vLLM Not published by the vendor
Next step up
Modal Example GPU list (per second) $0.001097
vLLM
Runs where
Modal
vLLM Web, desktop; hosted
Watch out
Modal $0 is the seat, not the GPU
vLLM The GPU is the bill
Last checked
Modal 2026-10-21 · modal.com/pricing
vLLM 2026-10-20 · github.com/vllm-project/vllm
Which one, then?
Pick vLLM if
platform and inference engineers who already have a GPU (or a kube cluster) and need an OpenAI-shaped endpoint.
Read the vLLM page