The latest articles on DEV Community by Ethan Mooneyham (@ethan_mooneyham_11d5c765a). https://dev.to/ethan_mooneyham_11d5c765a https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4223111%2F95a57d87-720d-4f0c-8359-96424133742a.png https://dev.to/ethan_mooneyham_11d5c765a en Ethan Mooneyham Tue, 03 Nov 2026 16:00:00 +0000 https://dev.to/ethan_mooneyham_11d5c765a/training-a-custom-model-without-renting-a-gpu-how-nistivo-quotes-trains-and-verifies-a-run-39i7 https://dev.to/ethan_mooneyham_11d5c765a/training-a-custom-model-without-renting-a-gpu-how-nistivo-quotes-trains-and-verifies-a-run-39i7 <p>Most custom model projects do not fail on the model. They fail on the plumbing around it: choosing hardware, writing a training configuration that actually fits that hardware, checking the dataset before spending anything, and finding out the real cost only when the run ends.</p> <p>Nistivo is a managed platform I built to take that plumbing away. You describe the model you need. The platform drafts a configuration, validates it against the hardware before quoting, and gives you one fixed price for the whole job. That price is final. There is no hourly meter.</p> <h2> How a run works </h2> <ol> <li> <strong>Describe the goal.</strong> Fine-tune a base model or train from scratch. Six families are supported: language, vision, image generation, speech, embeddings and classification.</li> <li> <strong>Configuration is drafted, then validated.</strong> The drafting step proposes architecture and training settings. Server-side validation checks them against real hardware limits: memory fit at the chosen context length, the platform's size ceilings, and a profit floor that keeps the quote honest. A configuration that does not fit is refused before it can be quoted, not after it fails.</li> <li> <strong>One fixed quote.</strong> The quote is computed from the validated configuration and locked. The GPU rate inside it is locked for 72 hours.</li> <li> <strong>Training with checkpoints.</strong> Checkpoints are saved and verified during the run. If a paid run fails before its first verified checkpoint, it is refunded in full. If it fails after one, it is re-run free from that checkpoint.</li> <li> <strong>Delivery with lineage.</strong> The finished model arrives as a versioned download: weights, configuration, checkpoints, and a model card covering what it was trained on, the recipe, intended use and limitations.</li> </ol> <h2> The tradeoff, stated plainly </h2> <p>A fixed quote means the job is scoped before it starts. If you want a shell on a GPU and full control of the training loop, a managed platform is the wrong tool, and this is a managed platform by design. What you get in exchange is a known price and a finished model you own outright.</p> <h2> Numbers </h2> <p>Runs use up to 8 GPUs at once, 16 on the Pro plan, across 15 GPU types. New accounts start with $100 in training credit, covering up to 40% of a quote. Pro is $49/mo for 20% off every quote.</p> <p>It launched today: <a href="https://nistivo.com/?utm_source=devto&amp;utm_medium=launch&amp;utm_campaign=launch-2026-11" rel="noopener noreferrer">https://nistivo.com/?utm_source=devto&amp;utm_medium=launch&amp;utm_campaign=launch-2026-11</a></p> <p>Happy to answer technical questions about the quoting math, the validation checks, or the checkpoint verification. What would you train first?</p> ai machinelearning devtools startup