> For the complete documentation index, see [llms.txt](https://docs.stg.ai.neevcloud.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.stg.ai.neevcloud.com/ai-inference/overview-3.md).

# Dedicated Inferencing

Run AI models on GPU hardware reserved exclusively for your deployment. No shared compute, no cold starts, no unpredictable latency.

## What is Dedicated Inferencing

Dedicated Inferencing lets you deploy AI models onto GPU hardware that belongs only to your workload. When you click **Deploy**, NeevCloud provisions a GPU instance configured for your chosen model. The GPU's VRAM, vCPU, and RAM are all reserved exclusively for that deployment — no other user or model shares it.

This is different from shared or serverless inference, where requests are routed to shared infrastructure. Dedicated Inferencing gives you a private, always-on endpoint that responds predictably, regardless of what other users on the platform are doing.

## Why use Dedicated Inferencing

Shared inference infrastructure introduces variable latency and unpredictable throughput as capacity fluctuates. Dedicated Inferencing eliminates these issues:

* **Consistent latency**: Requests always go to the same reserved GPU, not a shared pool that fluctuates under load.
* **Isolated VRAM**: The full VRAM of the configured GPU is available to your model with no memory pressure from other workloads.
* **No cold starts**: Your model is always loaded and warm. There is no startup delay between requests.
* **Hourly billing**: You pay for the running deployment, not per API call. Billing stops when you delete the deployment.

## Who should use this

Dedicated Inferencing is designed for:

* **AI/ML engineers**: Running production inference workloads that require a stable, long-running endpoint.
* **Platform teams**: Teams that need consistent latency guarantees under variable request loads.
* **Developers**: Deploying large models that require dedicated VRAM and cannot share GPU memory with other workloads.
* **Enterprises**: High-throughput or low-latency production requirements where shared infrastructure is not acceptable.

## Core Concepts

Understand the key ideas behind Dedicated Inferencing before you start deploying models.

### Dedicated vs Shared Inference

NeevCloud offers two modes of inference. Understanding the difference helps you pick the right option for your use case.

| Aspect             | Dedicated Inferencing                       | Shared / serverless                             |
| ------------------ | ------------------------------------------- | ----------------------------------------------- |
| **GPU allocation** | Reserved exclusively for your deployment    | Shared across multiple users and workloads      |
| **Latency**        | Consistent — GPU is always reserved for you | Variable — depends on shared load               |
| **Cold start**     | None — model stays loaded                   | Possible — model may unload between calls       |
| **VRAM**           | Full GPU VRAM available to your model       | Shared; affected by other workloads             |
| **Billing**        | Hourly per running deployment               | Per API call or per token                       |
| **Best for**       | Production workloads, low-latency apps      | Occasional use, bursty or unpredictable traffic |

### GPU Configuration

Each deployment runs on a GPU configuration — a combination of GPU type, GPU count, and VRAM. For validated models, NeevCloud pre-selects the recommended configuration, which you can change in the deployment drawer before deploying. For Bring Your Own Model deployments, you select your preferred configuration.

VRAM determines what fits: A model's parameter count maps directly to its VRAM requirement. A 7B model at FP16 needs roughly 14–16 GB of VRAM. A 70B model needs roughly 140 GB. Always check the VRAM field on a model card before deploying.

### Inference Endpoints

Once a deployment reaches the **Running** state, NeevCloud exposes a live HTTP endpoint for your model. You send inference requests to this URL and receive responses in real time. The endpoint stays live until you explicitly delete the deployment.

### Model Catalogue vs Bring Your Own Model

You have two ways to get a model running on Dedicated Inferencing:

* **Model Catalogue**: Browse NeevCloud's pre-validated open-source models. GPU configuration is recommended for you. Deploy with one click.
* **Bring Your Own Model**: Point to any HuggingFace repository. You choose the GPU configuration and pass custom serving arguments.