The agent era runs on your own models.

Today: frontier models

Most agents start on a frontier model through a provider's API. That's the right call for a prototype. The models are strong, the APIs are easy, and you can ship in a week.

Then the agent works, and usage grows. An agent doesn't call a model once. It plans, calls a tool, reads the result, and calls the model again, often dozens of times for one task. Every one of those calls carries the price and latency of a model built to do everything.

Teams also notice what they don't control: the model can change under them, prices can move, and their data leaves for someone else's servers.

Next: open models

Open models have closed most of the gap with frontier ones. Stanford’s 2025 AI Index found the gap between the top closed and top open-weight models on a leading benchmark shrank from 8% to 1.7% in a single year.¹ Teams that switch get lower cost, more control over latency, and a model that doesn't change unless they change it.

But an open general model is still general. It carries knowledge and skills the agent will never use, and it still has to be told, in every prompt, how this agent should work.

Then: your own model

An agent does a narrow set of things, over and over. That is exactly the setting where a small model, trained on those tasks, can match a far larger one. Researchers at NVIDIA argue that small language models are powerful enough for most agent tasks, and more economical.² Fine-tuned 7B models have outperformed GPT-4 on a wide range of specialized tasks.³

The training signal is already there. Every run of your agent produces a trace: what it was asked, which tools it called, what came back, whether the task succeeded. Train on those traces with supervised fine-tuning, then improve with reinforcement learning, where the model tries tasks in real environments and is scored on the result.

What you get is a model that is smaller, faster and cheaper to run, that you own, and that keeps improving as your agent does more work. Gartner predicts that by 2027, organizations will use small, task-specific models three times more than general-purpose ones, and that their own data will become a key differentiator.⁴

Your data is your playbook

There is a second reason to own the model: the data behind it. An agent’s traces record how your business actually works, step by step. Send them to someone else’s model and you hand over the playbook that makes your business hard to copy.

Keep the traces in your environment, train on them yourself, and the advantage they hold stays yours.

What it takes

Getting there takes more than a GPU cluster. You need somewhere to capture and clean traces, sandboxes where agents can act safely, thousands of them at once for reinforcement learning, a training stack, evals that compare the new model with the one you use today, and inference to serve it.

Usually those live with different vendors. Voltspan puts them on one platform, on NVIDIA clusters we build and run ourselves, with the sandboxes in the same data center as the GPUs.

How we train

We train on your tasks, your tool results and your feedback, starting from open models. We don't train on the outputs of closed frontier models. The models we train for you, and their weights, are yours.

Voltspan