On-Premise AI Platform

SynaptxCloud can build an on-premise AI platform around open-weight models that can be self-hosted. These include Llama (Meta), Mistral and Mixtral (Mistral AI), Falcon (TII UAE) and DeepSeek, as well as local fine-tuned versions of them. We design the system so that it is not tied to one model. The aim is that you can swap models without rebuilding your agent workflows. For classified environments, we can also work with proprietary models that a government has fine-tuned and cleared for on-premise use. Licence terms vary by model. We confirm them during project scoping.

It depends on the size of the model and how much traffic you expect. As a rough guide, these are typical figures. Entry-level, for 7B–13B parameter models: 2–4 NVIDIA A100 or H100 GPUs, suitable for 10–50 concurrent users. Mid-range, for 70B parameter models: 8–16 NVIDIA A100 or H100 GPUs, suitable for 50–200 concurrent users. Large-scale, for several 70B+ models: a custom GPU cluster, suitable for thousands of concurrent users. These figures are indicative. Real needs depend on your workload. SynaptxCloud gives hardware sizing guidance during project scoping.

Yes. SynaptxCloud can build an on-premise AI platform for air-gapped environments with no external network connection. In that design, model weights would be delivered physically on encrypted storage media and checked against cryptographic hashes. We build the system to run offline. Updates would arrive as verified patch packages, also delivered physically. Integrations would use internal APIs only. Air-gapped deployment is intended for organisations such as defence agencies, intelligence services and critical infrastructure operators.