> ## Documentation Index
> Fetch the complete documentation index at: https://docs.synapsai.cloud/llms.txt
> Use this file to discover all available pages before exploring further.

# Core concepts

> Readiness levels, pricing, credits, model lifecycle, and how SynapsAI Cloud manages compute and storage

SynapsAI Cloud hosts your models on [shared, multi-tenant GPU infrastructure](/isolation). You choose how models are kept warm, how they scale, and how you pay. This page explains the concepts that affect latency, availability, and cost.

## How SynapsAI Cloud works

1. You connect a [Hugging Face repository](/guides/hf-model-setup) and deploy it on the platform.
2. SynapsAI loads your model onto [shared GPU capacity](/isolation) and exposes an API endpoint scoped to your deployment.
3. You send inference requests with an [API key](/manage/api-keys) scoped to that deployment.
4. The platform [autoscales](/manage/autoscaling) instances based on traffic and tears down idle capacity according to your settings.

<Card title="Deploy your first model" icon="rocket" href="/quickstart">
  Follow the quickstart to go from account creation to your first API call.
</Card>

## Readiness levels

When you deploy a model, you choose a **readiness level**. This controls how quickly instances can serve traffic and how much you pay when idle.

| Level                          | Latency                                                          | Idle cost                                                | Best for                                           |
| ------------------------------ | ---------------------------------------------------------------- | -------------------------------------------------------- | -------------------------------------------------- |
| **Always ready**               | Lowest — at least one instance stays running                     | Highest — billed 24/7                                    | User-facing production APIs with low-latency needs |
| **Super fast** *(recommended)* | Low — instances load from prepped artifacts on demand            | Medium — billed only while instances are active          | Most production and staging workloads              |
| **Cold start**                 | Highest — weights are fetched from Hugging Face on first request | Lowest baseline — billed only while instances are active | Dev, testing, and infrequent batch jobs            |

<Note>
  **Always ready** includes storage for prepped artifacts at no extra charge. **Super fast** adds a [storage charge](#storage-for-super-fast-deployments) for those artifacts. **Cold start** does not use prepped storage.
</Note>

See [Deploy a model](/guides/deploy-model) for configuration details and [Optimizing costs](/guides/optimizing-costs) for tuning spend.

## Memory-based compute pricing

You are billed for the GPU memory (VRAM) consumed by active model instances.

**Example A:** One instance using 50 GB of VRAM costs approximately **\$5.50/hour** for that instance.

**Example B:** Two active instances, each using 50 GB, cost approximately **\$11.00/hour** in total.

Factors that affect memory usage:

* Model size and architecture
* Precision (for example, Float32 uses more memory than BFloat16)
* Quantization (INT4, INT8, FP8, EETQ)
* Context length for LLMs

All estimated costs are shown in the deployment UI before you confirm.

## Storage for Super fast deployments

When you choose **Super fast**, an additional storage charge applies for prepped model artifacts that enable rapid load times.

* **Price:** \$0.55 per GB per month
* **Example:** 20 GB of prepared artifacts → \$11.00/month
* **Example:** 60 GB of prepared artifacts → \$33.00/month

**Always ready** deployments include this storage at no extra charge.

## Credits and billing

SynapsAI Cloud uses a **credit** system. Credits map to US dollars and pay for compute, storage, and inference usage.

* Purchase credits from the [Billing](https://platform.synapsai.cloud/billing) page.
* Credits do not expire but are non-transferable and non-refundable.
* Enable auto-pay to avoid interruptions when your balance runs low.

See [Billing](/manage/billing) for purchasing credits, transaction history, and invoices.

## Model lifecycle

Every deployment moves through defined states. Track them on the model page in the dashboard or via the API.

| State        | Meaning                                                        | Can run inference?                                            |
| ------------ | -------------------------------------------------------------- | ------------------------------------------------------------- |
| **Building** | Weights are being fetched, validated, and optionally quantized | No                                                            |
| **Sleeping** | Deployed but no instance is loaded                             | Yes — the platform automatically loads an instance on request |
| **Ready**    | At least one instance is loaded and serving                    | Yes — lowest latency, no load delay                           |
| **Failed**   | Deployment or validation failed                                | No — check logs and redeploy                                  |

For operational details, see [Manage models](/manage/model).

## Autoscaling and worker timeout

Each model has its own autoscaling policy:

* **Scale up threshold** — average concurrent requests per instance before a new instance launches
* **Scale down threshold** — load level at which idle instances may terminate
* **Cooldown period** — minimum time between scaling events
* **Worker timeout** — idle time before the last instance shuts down (Super fast and Cold start only)

See [Autoscaling](/manage/autoscaling) for parameter guidance.

## Data handling

Inference requests are encrypted in transit (HTTPS), processed on [shared GPU infrastructure](/isolation) running your model, and not sold to third parties. SynapsAI does not intentionally retain request payloads beyond what is needed to complete processing.

For details on data location, retention, and video artifacts, see [Data](/resources/data).

## Supported tasks

SynapsAI Cloud supports a wide range of Hugging Face pipeline tasks — not only LLMs. Text generation, embeddings, image generation, speech, video, classification, and more are available depending on the model you deploy.

See the full [Supported tasks](/resources/tasks) table for pipeline names, streaming support, and API endpoints.