Overview
AMD has published a video on its YouTube channel about LegionEdge and the custom AI models we build. It’s embedded above.
A short video can only cover so much. This post fills in the background: why we build specialized models instead of defaulting to the largest general-purpose one, and what a custom model engagement with us produces.
Why custom models
Most teams start with the biggest general-purpose model they can afford, and for exploration that’s the right call. Production is different. Once a workload is defined, a smaller model trained on that workload’s own data can match a much larger one on the tasks that matter, at a fraction of the cost to serve.
The catch is that you have to define the workload first. Our decision framework for fine-tuning versus retrieval covers when a custom model is the right tool at all, and the field guide to distillation covers what a small model tends to lose when it’s trained carelessly, and how to keep it.
What an engagement produces
A custom model is more than a set of weights. Every model training engagement starts by choosing the task, the base-model candidates and the acceptance measures, and by establishing a baseline before any compute is spent on training.
At the end you hold four things: a dataset built for the task, with training and evaluation sets kept apart; the weights or adapters with their configurations and training logs; an evaluation harness you can rerun, with baseline comparisons and an error analysis; and a deployment handoff with inference examples, environment requirements and known limitations.
Build one with us
Thank you to the AMD team for telling this story. If you have a workload you’d like a model built for, scope a training run with us.
