Aeion AI — Fully On-Prem, Air-Gapped
Run AI in environments that cannot egress to OpenAI / Anthropic / Google. Ollama (or vLLM, HuggingFace TGI, Cerebras Inference) as the inference layer; Llama 3.1 / DeepSeek R1 / Qwen 2.5 / any open-weights model as the AI; full Aeion gateway running on YOUR network. HIPAA pre-BAA, FedRAMP High, DoD IL4-IL5, classified workloads, EU strict data residency.
When On-Prem AI Is Mandatory
Cloud AI providers cover most use cases — but not all. Five scenarios where on-prem is the only path:
Architecture — Air-Gapped Deployment
`
Inference-Layer Options
Aeion's ollama adapter is the front door but several inference engines speak the OpenAI-compatible API behind it.
Model Quality — Open Weights in 2026
Open-weights model quality is now competitive with cloud frontier models for most workloads.
Hardware Sizing
Three reference deployments.
FAQ
No. Air-gapped deployments — available to any tenant, no separate tier required — have zero outbound network calls. No license-server check, no telemetry, no model-update phone-home. The license is a cryptographically-signed file installed at deploy time.
Manual model pulls on a sneakernet-style update cadence. Pull new Llama / DeepSeek / Qwen releases to a staging environment with internet; physically transfer + install into the air-gapped environment. Most regulated deployments do this quarterly.
Yes — hybrid is supported. Mark sensitive collections / agents to use only `ollama:*` models; let general workloads route freely. Aeion's per-tenant routing rules enforce.
Multiple options: sentence-transformers via TEI (Text Embeddings Inference), Nomic embed-text on Ollama, Mixedbread self-hosted. All speak OpenAI-compatible API; Aeion's `embed()` works against any.
Whisper on Ollama (or whisper.cpp directly) for STT; Kokoro on local GPU for TTS; Coqui XTTS-v2 for voice cloning. Quality is slightly behind cloud (ElevenLabs / Cartesia) but acceptable for most workloads.
Stable Diffusion (SDXL, SD3, Flux) on local GPU for image — quality is competitive with cloud Stability/Ideogram. Video gen on-prem is harder; open-weights video models lag cloud Luma/Runway significantly. For classified workloads needing video gen, plan around the gap.
One additional GPU server to operate (or cluster). Aeion's deployment + monitoring tools cover it; ops complexity is roughly equivalent to running one more Postgres instance.