On-premise and localized model deployment
When inference cannot leave your network, the hard part is everything around the model: accelerator support, engine tuning, monitoring and upgrades. That is the part we deliver.
5 vendors
Accelerators
Air-gap capable
Network
LLMFabric
Control plane
Apache 2.0
Licence
Your models, your hardware, your network boundary
Hardware and capacity sizing against your real concurrency and latency targets
Installation inside your network, air-gapped where required, with no callback to us for the system to run
Accelerator support beyond NVIDIA: AMD GPU, Huawei Ascend NPU, Hygon DCU and MThreads GPU
LLMFabric as the control plane: open source under Apache 2.0, so you are not locked into us
Grafana and Prometheus dashboards, plus operator handover documentation and training
Before you ask
- Can the deployment run fully air-gapped?
- Yes. The control plane, inference engines and model weights are installed inside your network with no outbound dependency required at runtime. Updates are delivered as verified artifacts you apply on your own schedule.
- Do you support non-NVIDIA accelerators?
- LLMFabric supports NVIDIA GPU, AMD GPU, Huawei Ascend NPU, Hygon DCU and MThreads GPU. This matters for teams with domestic hardware mandates or existing non-NVIDIA estates.
- What happens if we stop working with you?
- LLMFabric is open source under Apache 2.0 and the deployment runs on your hardware. You keep a working system and can operate it yourself; the handover documentation is written on that assumption.
Other workloads
AI Training
Dense GPU nodes with high-bandwidth interconnect, wired for distributed training runs that cannot afford to restart from scratch.
AI Inference
Production inference on vLLM, SGLang or TensorRT-LLM, metered per token and reachable through a single API surface.
Fine-tuning
LoRA, QLoRA and full-parameter fine-tuning, followed by the quantization and evaluation work that makes the result deployable.