# A sovereign AI platform on your GPUs.

AVELIN's inference layer installs on your GPUs and is optimized for them at GPU level. It runs nine AVELIN models, each fine-tuned for one kind of work, with an API platform on top. You decide where it runs and who can reach it.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)
- Try the live demo (https://arena.avelin.ai)

Figure. The sovereign platform, from your apps down to your GPUs.
Alt text: Layer diagram of the AVELIN sovereign platform. From top to bottom: your apps and agents; the API platform; the Decision Layer; the nine AVELIN models; the AVELIN inference layer; and your GPUs. A frame around everything except your apps is labelled "Runs on hardware you control".
- API platform: OpenAI- and Anthropic-compatible API · speech · web search and scraping · cost control · user management
- Decision Layer: Optional. Picks the level of model for each request.
- Nine AVELIN models: Chat · Coding · Agents, each in fast, pro and ultra
- Inference layer: Optimized for your GPUs and the nine models
- Your GPUs

### What is inside the platform

Three parts, one stack. It runs on hardware you control, in your network.

#### Inference layer, optimized for your GPUs

Inference is the step where a model reads a prompt and writes an answer. The AVELIN inference layer runs that step on your GPUs, optimized at GPU level for the nine AVELIN models, so your hardware generates tokens efficiently.

#### 3 families · 9 models

Each model is fine-tuned for one type of task. Chat covers knowledge work and is mostly multimodal: images, documents and PDFs. Coding covers math and code, with large context windows. Agents covers volume work that runs 24/7, built for speed and price, with a reduced context. Each family has fast, pro and ultra tiers.

Figure. The nine AVELIN models. Context sizes are those of AVELIN Cloud.
Alt text: The nine AVELIN models: three families, each in fast, pro and ultra
| Family | What it is for | Context | fast | pro | ultra |
| --- | --- | --- | --- | --- | --- |
| Chat | Knowledge work. Mostly multimodal: images, documents and PDFs. | 256K | avelin-fast | avelin-pro | avelin-ultra |
| Coding | Math and code. Large context windows for the most complex tasks. | 1M | avelin-coding-fast | avelin-coding-pro | avelin-coding-ultra |
| Agents | Volume work that runs 24/7, built for speed and price. Handles complex tasks, including coding, with a smaller context. | 256K | avelin-agentic-fast | avelin-agentic-pro | avelin-agentic-ultra |
- fast: Speed and low cost.
- pro: Balanced, for medium complexity.
- ultra: The hardest tasks, where intelligence matters.

- See all nine models (https://avelin.ai/models)

#### Cost

On AVELIN Cloud, tasks run at up to 90% lower cost than flagship models at list prices, measured cost only in our live Arena. Self-hosted goes further. There is no per-token bill: your marginal cost is your own hardware, and prices are set by agreement. A proof of concept on your workloads shows the real number.

#### API platform with cost control

Your apps call one API. It works with OpenAI and Anthropic clients: change the base URL and the API key.

- Chat API, compatible with OpenAI and Anthropic clients
- Speech-to-text and text-to-speech
- Web search and web scraping
- Observability and cost control: cost by team, key and model, with budgets and caps
- User management and scoped API keys

- See the API platform (https://avelin.ai/api-platform)

### Make the models fit your workflows

Enterprises can have the models fine-tuned or trained on their own workflows. The aim is a model that fits your tasks, not a generic one.

The written scope states the workflow, the data we may use and how success is measured.

- Request a POC for tuning (https://avelin.ai/poc?topic=enterprise#request)

### Tuned to work with the Decision Layer

The nine models are tuned to work with the Decision Layer. For each request it picks the model the task needs: a cheaper one when the task is easy, a more capable one when it is hard.

You can use the models without it. You can also start with a POC of the Decision Layer, then move to the platform.

Figure. Prove, move, own: the path from a POC to your own GPUs.
Alt text: Three steps. Prove: a POC of the Decision Layer on the providers you use today. Move: AVELIN Cloud, with the nine models and the Decision Layer. Own: the platform on your GPUs, as an enterprise or hybrid deployment.
1. Prove: A POC of the Decision Layer on the providers you use today.
2. Move: AVELIN Cloud: the nine models and the Decision Layer.
3. Own: The platform on your GPUs, as an enterprise or hybrid deployment.

- How the Decision Layer works (https://avelin.ai/decision-layer)

### Three ways to deploy

Choose where the models run.

- AVELIN Cloud The nine models and the Decision Layer, run by AVELIN, with inference in the United States, Europe and MENA. Connect your own providers to cut token spend. Isolated in-country locations are planned.
- Enterprise The Decision Layer, the full platform, or both, on your hardware. Optimized for your GPUs, with models fine-tuned to your workflows.
- Hybrid Local models on your GPUs, with offload to a public model when a task needs more intelligence or the local model is under heavy load. A privacy check lets you decide whether detected private data may leave.

- Compare the three deployments (https://avelin.ai/deployment)

### Privacy, at three levels

Privacy is three questions. What do you keep? Where does it run? Can it leave?

Three privacy levels. You choose the level that fits the workload.
| Level | What happens to your prompts |
| --- | --- |
| AVELIN Cloud | Zero retention by default. Not used to improve AVELIN. Opt in to share anonymized requests to help improve AVELIN. |
| Hybrid | A privacy check lets you decide whether detected private data may go to a public model or must stay on your GPUs. |
| On-premise | The ultimate privacy: prompts are processed on your own GPUs. |

- How privacy works (https://avelin.ai/privacy)
- Read the data policy (https://avelin.ai/legal/data-policy)

### How a deployment starts

1. Prove it with a POC A POC, or proof of concept, is a short test on one of your workloads. We test the Decision Layer beside your current setup and report cost per task.
2. Scope it in writing The scope covers where it runs, what leaves, who has access, hardware, updates, licence, support, time and price. It lists every network path the software opens.
3. Install it on your GPUs The platform is installed on hardware you control, as agreed in the scope.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

- What a scope settles (https://avelin.ai/private-deployment)

### FAQ: Questions about the sovereign platform

**What is a sovereign AI platform?**
A sovereign AI platform means you, not a vendor or a cloud, decide where the system runs and who can reach it. With AVELIN, the platform runs on GPUs you control, with your data and your prompts inside your network. That is an on-premise LLM stack with data residency you control.

**What is included?**
Three parts: the AVELIN inference layer that runs the models on your GPUs, the nine AVELIN models, and the API platform. The API platform adds an API compatible with OpenAI and Anthropic clients, speech-to-text, text-to-speech, web search, web scraping, an observability and cost-control platform and user management. The Decision Layer can sit on top.
More: The API platform (https://avelin.ai/api-platform)

**How is it optimized for our GPUs?**
The AVELIN inference layer is optimized at GPU level for the nine AVELIN models, so your hardware generates tokens efficiently. The GPU class and the number of GPUs depend on the tiers you run, and the written scope states both.

**Can the models be tuned to our workflows?**
Yes. For enterprise customers the models can be fine-tuned or trained on specific workflows. The written scope states the workflow, the data used and how success is measured.

**Do we need the Decision Layer?**
No. You can use the nine models on their own. They are tuned to work with the Decision Layer, which picks the model each request needs, and a POC of the Decision Layer is a common way to start.
More: Decision Layer (https://avelin.ai/decision-layer)

**Who can access a deployment on our GPUs?**
You decide who can reach it. The written scope lists every network path the software opens and whether each is on or off, and how any vendor support access is granted.
More: What a scope settles (https://avelin.ai/private-deployment)

**How is it priced?**
A deployment on your GPUs is priced by agreement in the written scope. In AVELIN Cloud the nine models have public list prices, from $1.14 to $12.50 per 1M output tokens.
More: Pricing (https://avelin.ai/pricing)

**How does a deployment start?**
With a POC on one workload, then a written scope, then the install on your GPUs.
More: What a scope settles (https://avelin.ai/private-deployment)

Related: Decision Layer (https://avelin.ai/decision-layer); AVELIN models (https://avelin.ai/models); API platform (https://avelin.ai/api-platform); Deployment (https://avelin.ai/deployment); Privacy (https://avelin.ai/privacy); Private deployment scope (https://avelin.ai/private-deployment)

### Start with a POC.

Tell us one workload. We test the Decision Layer on it and report the cost per task. Then we scope the platform on your GPUs.

- Request a POC (https://avelin.ai/poc?topic=enterprise#request)

- Try the live demo (https://arena.avelin.ai)

- Start building (https://avelin.ai/docs/api/quickstart)

URL: https://avelin.ai/sovereign-ai-platform
Updated: 3 Oct 2026