starus.io
Status: online

Private AI cloud

Inference that never leaves hardware we own.

A GPU-backed private cloud on a self-hosted six-node k3s cluster. Chat, transcription, speech, images and video are served by local models behind one OpenAI-compatible gateway — no third-party inference API, and nothing sent to a vendor for processing.

  • OpenAI-compatible API
  • Open-weight models
  • Declared in git
  • Zero Trust access
6
k3s nodes
one cluster, one LAN
5
GPUs
3x RTX 4090 + 2x RTX 2080
~470 GB
RAM
aggregate across the fleet
170+
CPU threads
schedulable

The interface

One endpoint, every modality.

A single LiteLLM gateway fronts the whole fleet. Point an existing SDK at it and change the base URL — the same code that called a vendor now calls the rack, with one key and one request format across text, audio, image and video.

same client, local hardware
from openai import OpenAI

client = OpenAI(
  base_url="https://<gateway>.starus.io/v1",
  api_key=STARUS_KEY,
)

client.chat.completions.create(...)
client.audio.transcriptions.create(...)
client.audio.speech.create(...)
client.images.generate(...)
# video generation on the same key
Illustrative. Gateway hostnames and keys are issued privately; nothing here is a public endpoint.

  • Chat & vision

    Text and image-in conversations against a locally served model.

  • Speech to text

    Batch and file transcription, many times faster than realtime.

  • Text to speech

    Low-latency synthesis running on the same fleet.

  • Image generation

    Text to image, plus instruction-driven editing of existing frames.

  • Video generation

    Text to video, with a restoration pass for upscale and cleanup.

  • Image understanding

    Captioning, OCR and object detection as ordinary API calls.

Models in service

Open weights, loaded locally.

Every model runs from local storage on local GPUs. Throughput notes are what this hardware does, not vendor figures.

  • Gemma 4 E4BChat, visionserved via llama.cpp
  • Whisper large-v3Speech to text~9x realtime
  • SupertonicText to speech~25x realtime
  • Florence-2Caption, OCR, detectionimage understanding
  • Z-Image TurboText to imagefast sampler
  • Qwen-Image-EditImage editinginstruction-driven
  • LTX-2Text to videodiffusion video
  • SeedVR2Video restorationupscale and cleanup

The fleet

Five GPUs across six nodes.

One k3s cluster on a single LAN. Workloads land on whichever node has the VRAM, and the heavy diffusion and video jobs get the 4090s.

  • gpu-01

    RTX 4090

    24 GB VRAM

  • gpu-02

    RTX 4090

    24 GB VRAM

  • gpu-03

    RTX 4090

    24 GB VRAM

  • gpu-04

    RTX 2080 + 2080 SUPER

    8 GB each VRAM

Plus two CPU-only nodes carrying the platform services.

Platform

Deliberately boring underneath.

The models are the interesting part. Everything below them is conventional on purpose: declarative, observable, and reproducible from a git repository.

  • ArgoCD

    GitOps. Every workload is declared in git and reconciled.

  • Infisical

    Secrets management, out of the manifests.

  • Prometheus + Grafana

    Metrics, dashboards and alerting for the fleet.

  • Loki

    Log aggregation across every namespace.

  • MinIO

    S3-compatible object storage for models and artifacts.

  • Supabase

    Postgres, auth and storage for applications.

  • Odoo

    ERP running on the same cluster as the inference stack.

  • Cloudflare Zero Trust

    Identity-aware access in front of every service.

Access is controlled.

Every service on the cluster sits behind Cloudflare Zero Trust. Subdomains are not open to the internet — a request is checked against an identity policy before it ever reaches a pod. There is nothing to sign up for on this page.