Run frontier AI entirely
on your own hardware.

Cube AI is a self-hosted platform for running large language models privately. It bundles inference, retrieval-augmented generation, guardrails, governance, audit, and a production UI into one stack that runs entirely on infrastructure you control.

With a hosted API, your prompts and data leave your organization and are processed in someone else's cloud and jurisdiction. Cube AI runs inside your own perimeter — on-prem, air-gapped, or in a sovereign cloud — so data never leaves, and you own the models, the policies, and the audit trail.

Cube AI is model-agnostic. It serves open-weight models such as Llama, Mistral, Qwen, DeepSeek, Phi, and Gemma — plus custom fine-tunes in GGUF or safetensors format — through vLLM, Ollama, and Hugging Face, all from a single control plane.

Anywhere your data has to live: on-premises behind your firewall, fully air-gapped for classified environments, in an EU or national sovereign cloud, in your own private VPC, or at the edge — with the same governance everywhere.

Guardrails are policy rules that sit between your users and the model — inspecting every prompt and response before it is processed or returned. They can block harmful inputs, detect prompt-injection attempts, filter profanity or off-topic requests, and enforce domain-specific rules your organization defines.

Cube AI uses NVIDIA NeMo Guardrails as its policy engine, augmented with Microsoft Presidio for automatic PII detection and redaction. You author guardrail configurations as YAML-based Colang rules through the Cube dashboard — no code required. Rules hot-reload with zero downtime, and every guardrail decision is recorded in the audit trail so you can see exactly what was allowed, blocked, or redacted on every call.

A TEE is a hardware-isolated region inside a CPU — such as an AMD SEV-SNP confidential VM or an Intel TDX Trust Domain — where code and data are encrypted in memory at all times. Even the hypervisor and the host OS cannot read the contents. Remote attestation lets you cryptographically verify that the correct, unmodified software is running inside the enclave before trusting it with sensitive data.

Deploy Cube AI on a host with AMD SEV-SNP or Intel TDX hardware — available on several cloud providers and on-prem servers. Cube AI runs on top of Cocos AI, Ultraviolet's open-source confidential-computing layer, which handles enclave provisioning, remote attestation, and key management automatically. Once running inside a TEE, every model weight, prompt, and response is encrypted in use, and clients can verify the enclave before sending any data.

Yes. The core is Apache 2.0. You can inspect the source, fork it, and run it indefinitely with no vendor lock-in. Commercial support and enterprise features are available from Ultraviolet.