A frontier model is expensive to train, expensive to serve, and too blunt to be the unit of iteration.
That is the quiet premise behind MinT, Mind Lab's new managed infrastructure system for training and serving large numbers of LoRA-adapted LLM policies. The headline number is the fun part: million-scale policy catalogs over shared 1T-class base models. But the deeper point is less about scale theater and more about where AI engineering is moving.
The paper's bet is simple. Keep the base model resident. Move the adapters.
That sounds like plumbing because it is plumbing. Rollout, update, export, evaluation, serving, rollback. The unglamorous verbs. But those verbs are increasingly the product surface. If every customer, task, workflow, or agent needs its own tuned behavior, you cannot ship merged checkpoints like they are application releases. You need policy addressability. You need rollback. You need cold-loading treated as scheduled service work. You need the adapter to become the deployable artifact.
This is the part of the AI stack that still gets underpriced. We talk about model capability as if it arrives in named releases: bigger context, better reasoning, higher scores. But the operational question is different. How many distinct behaviors can you safely train, evaluate, serve, and retire without turning the system into a warehouse of brittle checkpoints?
MinT's answer is to separate the expensive base from the fast-moving policy layer. Adapter-only handoff, concurrent GRPO, packed MoE LoRA tensors, million-scale catalogs. The vocabulary is technical, but the shape is familiar: software ate the model release.
The obvious counterpoint is that LoRA infrastructure is not intelligence. A bad policy catalog is still bad. More adapters can mean more surface area, more evaluation debt, more ways to route the wrong behavior into the wrong context.
True. But that is exactly why the infrastructure matters. Once models become shared substrates, intelligence is no longer only in the weights. It is in the system that decides which behavior exists, when it is active, how it is measured, and how quickly it can be replaced.
The next AI platform may not look like a smarter chatbot. It may look like a control plane.