0.1 / how SPLIT works

An AI model is a sequence of layers. Each layer transforms information from the one before it. Running the whole model on one server can require more memory and compute than your company's hardware budget allows.

SPLIT runs the model across two environments. A small set of layers runs on your hardware. The compute-intensive part runs on our GPU servers. Both exchange intermediate results to complete inference and generate a response to your API request.

0.2 / one model, two environments

The split depends on your model, hardware and workload. Your configuration determines which layers stay in your environment and how much compute runs on SPLIT GPUs.

0.3 / APIs built for your needs

Personalize the layers on your hardware.

Safe AI starts with the first N local layers and client-side adaptation. Steering uses selected middle layers. Local LoRA or selective fine-tuning specializes final layers; QLoRA can reduce training memory on compatible hardware. Decision heads extract task-specific scores and predictions from model representations.

Local access control selects the adapters and knowledge collections each authenticated user is allowed to use, on your server or a managed PC. The policy service checks every request before retrieval and adapter selection. Permissions come from verified identity, never from the prompt. This adds no Transformer layers; the selected adapters still need their host layers locally.

Layer-based techniques need the right local positions. A middle or final block can enable additional customization, subject to model-specific runtime validation. More local blocks add weights and compute. Each separate local block also requires another exchange with the remote partition. Access control instead adds a local policy service, without increasing the layer count.