Our inference
The Grom lineup
The full cycle stays in-house: we research architectures, train and fine-tune models on our own data, and run inference on our own hardware. That gives us control over quality, latency and the cost of every generation.
Infrastructure
Our own hardware, our own perimeter
Grom models run on our own servers: we operate the racks, the network and the request queue ourselves, so speed and stability are on us — not on a third-party provider.
The inference fleet is built around NVIDIA H200 and A100, A5000 and other accelerators: every task runs on the hardware that suits it, from fast text answers to video generation and 100-megapixel upscaling.
There are two ways to work with the models: in the GPTunneL interface, like any other model in the chat, or over the API — with the same key that gives access to the other two hundred neural networks, no separate integration required.
For business we agree on a higher SLA on request: dedicated capacity, guaranteed response time and a priority queue. When security requirements are stricter, Grom models can be deployed on-premise, inside the company's own perimeter.