Modular (Mojo/MAX) review

GPU work is metered per minute on Modular (Mojo/MAX), and the service also bills on the hourly and free tier model. Short jobs therefore stop at the minute rather than rounding to a full hour, which suits training runs and inference endpoints that idle between bursts.

On the hourly model, input prices run $0.10 to $1.74 and output prices $0.25 to $4.40. Those entry points sit above the €0.04 median and the €0.14 median recorded across the GPU providers on the site, so expect a higher token bill than at the cheapest listings.

The GPU line-up covers NVIDIA B200, AMD MI355X and Google TPU. That spread matters if you want to weigh a NVIDIA part against an AMD, TPU, Trainium or Qualcomm option inside the same account, rather than moving between separate vendor catalogues.

New Modular Cloud accounts come with credits, so a first API call or a short GPU run can start before any spend. Billing runs on the hourly and free tier model, and the OpenAI-compatible API means those early calls use client code you already have.

The largest context window in the catalogue is 10,000,000, against a 1,000,000 median across the sites listed. That gap suits long documents and codebases in a single request, though the window attached to a specific model is what your code sees.

The catalogue covers GLM 5.2, Gemma 4 31B and MiniMax M3, spanning general chat, coding, and image and video models. Names are written as they appear in the model list, so you can compare them directly with the OpenAI-compatible client you already run.

Modular Inc is a US company listed under US, and its compliance certifications cover SOC 2 Type 2. Buyers who need a formal audit trail for their own customers should check that the certificate covers the products and regions they buy.

Support runs on email, chat and help centre. Email, chat and a help centre cover most questions, with a community channel for shared problems, but phone is not among the listed routes, so urgent issues need a different approach.