oneStack
Every open model on your own GPUs. One API in front of all of them.
oneStack puts open-weight models from the Hugging Face catalog onto your NVIDIA servers, for chat, vision, embeddings, search, speech and images. It slices the cards, routes the traffic and bills each team for what it uses.
- Whole cards or MIG slices
- Each deployment gets a full GPU or an isolated slice of one, with its own network per organisation.
- The Hugging Face catalog
- gpt-oss, Kolibri, Llama, Qwen, Mistral and thousands more, searched and deployed from the console.
- OpenAI-compatible
- Existing SDKs point at your gateway and keep working.
What it does
From GPU to API, run by one console. Deploy, slice, route and bill without writing a compose file.
One wizard, eight engines
Pick an engine, a model and a card; oneStack writes the deployment, routes it, issues the certificate and checks its health.
Slice a card per tenant
NVIDIA MIG splits one GPU into isolated slices, and each organisation gets its own network on top.
Routing that uses the cache
Requests that share a prompt prefix land on the same instance, so its cache hits; a failing instance is taken out and traffic moves on.
Live users go first
Interactive requests overtake bulk jobs on the same GPUs, so indexing never makes a chat wait.
Know what fits before you deploy
The calculator adds weights, KV cache and concurrency, and recommends the slice that holds them.
Cost per team
Each instance costs its share of the server's VRAM, rolled up per organisation and team, ready to export.
Models
The open-weight models people know. And thousands more from Hugging Face.
Examples. Search the whole Hugging Face catalog from the console, or upload your own weights.
In daily work
Your models, used like any other API.
index.tsconst gpu = new OpenAI({ baseURL: "https://ai.nordwerk.internal/v1", apiKey: process.env.GATEWAY_KEY,});await gpu.chat.completions.create({ model: "gpt-oss-120b"On your GPUs, messages,});Point the SDK, keep the code
Anything written for OpenAI's API talks to your gateway: chat, responses, embeddings, rerank and audio.
Engine image updated06:10
vLLM on three instances in group chat. Health checks passed.
Rolled back06:14
Instance 4 failed its health check and runs the previous image again.
Updates that undo themselves
A new engine image is rolled onto the instances with a health check, and an instance that fails it goes back to the image before.
Embeddings · RTX PRO 6000 BlackwellBenchmark61,206tokens per second
An open-weight embedding model on two cards, 2.0 times the rate of one.
- One card30,544
- Two cards61,206
Measured, not promised
Two Blackwell cards serve an open-weight embedding model at twice the rate of one.
Workloads
What runs on it. Five kinds of work, on your own GPUs.
- ModelOpen-weight embeddings
- Throughput on two cards61,206 tok/s
- Large batchesSplit across instances
- StatusHealthy
Documents and messages are embedded on your own GPUs, so search never leaves your infrastructure.
Native in oneAI
Your models, inside oneAI. Private models next to the frontier models, under the same rules.
oneAISearch, reading, compliance- Private model01
- LoRA adapter04
- interactive · background03
Everything in oneStack
Built for a GPU fleet, not a demo. Everything to run open models on your own GPUs.
Models and engines
- vLLM for chat, vision, embeddings, reranking and speech
- TEI for embeddings, Ollama for GGUF models
- ComfyUI for images, Speaches for speech to text
- The Hugging Face catalog, searchable from the console
- Ollama models and your own uploads
- A signed, digest-pinned model registry
- FP8, AWQ, GPTQ and INT8 weights, FP8 KV cache
- Tool calling, reasoning, speculative decoding, prefix caching
- Up to four LoRA adapters per instance
GPUs
- NVIDIA, a whole card or a MIG slice per instance
- MIG on H100, H200, A100, A30 and RTX PRO 6000 Blackwell
- Slices restored after a reboot
- VRAM calculator with KV cache and concurrency
- Slice recommender for each model
- Choose the card a deployment lands on
Gateway
- Chat completions, responses, embeddings and models
- Rerank and score, Cohere-compatible
- Speech, transcription and translation
- Availability groups across servers behind one model name
- Round robin, least connections, weighted and prefix-hash routing
- Circuit breaker, failover and stall detection
- Priority scheduling, batch splitting, aliases
- Keys with rate and token limits, team scope and expiry
Tenants and cost
- Organisations, teams and users with roles
- VRAM quotas per organisation and team
- Model permissions per organisation
- A network of its own per organisation
- Cost per instance from its share of VRAM
- Reports per organisation and team, with export
- Two-factor sign-in and an audit log
Operations
- Templates that deploy to many servers at once
- Quota check before a batch deploy, all or nothing
- Redeploy with health wait and automatic rollback
- Drift check between stored and running configuration
- 13 Grafana dashboards and 19 alert rules
- Mutual TLS between console and servers
- Root operations through a separate, allow-listed helper
- Local models keep running without internet
Install
- A Debian package and systemd on each GPU server
- Automatic TLS through Traefik and Let's Encrypt
- Backup and restore of each server's setup
- On your hardware, or GPU servers through a partner
Questions
What platform teams ask first.
Another question?Talk to us
The oneAI family
Three building blocks, one sovereign AI OS. Sovereignty is a matter of freedom of choice — we offer the choice.
oneAI, oneStack and oneDeploy run in the cloud, in your private cloud or on-premise. Each works on its own, and all of them work together natively.
Next step
Put your GPUs to work.
Thirty minutes with a solutions engineer, your hardware and the models you want to run.
Talk to a partnerBook a demo
Thank you.
Your request is with us. A person from oneAI answers within one business day.
A person from oneAI answers within one business day. No ticket queue. By sending the form you agree that oneAI stores and processes your details to answer your request. Privacy policy