# Prime Intellect

> Train, evaluate and deploy custom AI models with integrated training, inference, GPU compute and sandboxes.

## Product overviews in Markdown

- [Homepage](https://www.primeintellect.ai/.md)
- [Model Factory](https://www.primeintellect.ai/model-factory.md)
- [Compute and Sandboxes](https://www.primeintellect.ai/compute.md)
- [Inference](https://www.primeintellect.ai/inference.md)

## Get started

1. Create a [Prime Intellect account](https://app.primeintellect.ai) and choose your personal or team workspace.
2. Install [uv](https://docs.astral.sh/uv/getting-started/installation/), then install and authenticate the CLI:

```bash
uv tool install prime
prime login
prime config view
```

Update an existing installation with `uv tool upgrade prime`. Use `prime --help` and each subcommand's `--help` for your installed version. See the [CLI reference](https://docs.primeintellect.ai/cli-reference/introduction.md) and [configuration guide](https://docs.primeintellect.ai/cli-reference/config-cli.md).

3. Create a scoped, expiring [API key](https://docs.primeintellect.ai/api-reference/api-keys.md) in [Settings → API Keys](https://app.primeintellect.ai/dashboard/tokens). Grant only the product permissions you need. Store credentials in server-side environment variables or a secret manager, never browser code, public repositories or trace content.
4. Confirm the workspace and billing account before creating resources. [Team configuration](https://docs.primeintellect.ai/cli-reference/teams.md) and [billing dashboard](https://app.primeintellect.ai/dashboard/billing) help you check ownership and credit balance.

CLI team configuration does not automatically select a team for direct inference HTTP/SDK requests: include `X-Prime-Team-ID` for team inference billing. Traces use their own documented workspace selection, including `PRIME_TEAM_ID`. Do not assume all products use the same team header or environment variable.

## Choose your first workload

| Goal | Start here | What you need |
| --- | --- | --- |
| Call a model | [Inference quickstart](https://docs.primeintellect.ai/inference/overview.md) | Inference permission, funded account, a current model ID |
| Train a specialist model | [Hosted Training quickstart](https://docs.primeintellect.ai/hosted-training/getting-started.md) | Beta access, Python 3.10+, funded wallet, environment/config and checkpoint storage |
| Test a model or agent | [Hosted evaluations](https://docs.primeintellect.ai/tutorials-environments/hosted-evaluations.md) | An environment and model endpoint |
| Build an RL environment | [Environment quickstart](https://docs.primeintellect.ai/tutorials-environments/getting-started.md) | CLI login and a profile username |
| Run isolated code | [Sandbox quickstart](https://docs.primeintellect.ai/sandboxes/overview.md) | Sandbox permissions and a container image |
| Rent GPUs | [GPU provisioning guide](https://docs.primeintellect.ai/cli-reference/provision-gpu.md) | Current availability, billing and SSH configuration |
| Run your own training stack | [prime-rl](https://docs.primeintellect.ai/prime-rl/overview.md) | Compatible GPU hardware and framework configuration |
| Inspect training or agent traces | [Prime Traces](https://docs.primeintellect.ai/traces/overview.md) | Traces permissions, correct workspace and supported records |

## Documentation for agents

Read the [platform overview](https://docs.primeintellect.ai/platform-overview.md) for how the products fit together. Fetch the [complete documentation index](https://docs.primeintellect.ai/llms.txt) to discover current guides and API endpoints. Documentation pages are available as Markdown by appending `.md`; the [full documentation export](https://docs.primeintellect.ai/llms-full.txt) is available when you need the whole reference, but the index and task-specific pages are usually more efficient.

These website pages summarize products and route you to implementation guides. Use the live docs, model catalog and availability APIs for current schemas, supported models, access requirements, quotas and prices. Do not infer API endpoints from marketing-page URLs.


## Training, evaluations and environments: first workflow

Hosted Training is currently beta: [request access](mailto:contact@primeintellect.ai) before dispatching. For dedicated Model Factory clusters, [contact our team](/contact). For self-hosted training, use [prime-rl](https://docs.primeintellect.ai/prime-rl/overview.md).

1. Define the task and measurable reward. Read the [environment model](https://docs.primeintellect.ai/hosted-training/environment-model.md), then [create an environment](https://docs.primeintellect.ai/tutorials-environments/create.md) or [install one from the Hub](https://docs.primeintellect.ai/tutorials-environments/install.md).
2. Evaluate your baseline with [local evaluations](https://docs.primeintellect.ai/tutorials-environments/evaluating.md) or [hosted evaluations](https://docs.primeintellect.ai/tutorials-environments/hosted-evaluations.md). Keep evaluation separate from the training data.
3. Follow the [complete training walkthrough](https://docs.primeintellect.ai/hosted-training/end-to-end-run.md). Inspect available models/capacity with `prime train models`; use the documented TOML configuration, not a guessed dispatch schema.
4. Configure a [volume](https://docs.primeintellect.ai/hosted-training/volumes.md) and a `[ckpt]` block before launching if you need model weights. **Checkpointing is off by default; without checkpoint storage, a completed run can leave metrics but no saved model.** Check [volume sizing](https://docs.primeintellect.ai/hosted-training/volume-capacity.md).
5. Monitor with `prime train logs <run-id> -f`. Use `prime train stop <run-id>` to stop early. Review reward curves, rubric scores and traces for reward hacking, then evaluate the saved model.
6. Wait for final volume sync, download checkpoints and follow the [HF export instructions](https://docs.primeintellect.ai/hosted-training/end-to-end-run.md#export-to-hf-format) before serving DCP-sharded weights.

### Training reference

- [Hosted RL configuration and lifecycle](https://docs.primeintellect.ai/hosted-training/hosted-rl.md)
- [Hosted SFT](https://docs.primeintellect.ai/hosted-training/hosted-sft.md)
- [Supported models](https://docs.primeintellect.ai/hosted-training/models.md)
- [Billing](https://docs.primeintellect.ai/hosted-training/billing.md)
- [Pause and resume](https://docs.primeintellect.ai/hosted-training/pause-resume.md)
- [Troubleshooting](https://docs.primeintellect.ai/hosted-training/troubleshooting.md)
- [Practical training recipes](https://docs.primeintellect.ai/guides/recipes.md)
- [Search-agent training](https://docs.primeintellect.ai/guides/search-agents.md)
- [Browser environments](https://docs.primeintellect.ai/guides/browser-environments.md)
- [Environment secrets](https://docs.primeintellect.ai/tutorials-environments/secrets.md) and [environment variables](https://docs.primeintellect.ai/tutorials-environments/environment-variables.md)

### Trace storage and observability

Use [Prime Traces](https://docs.primeintellect.ai/traces/overview.md) to inspect, search, download and query supported traces. Follow the [upload guide](https://docs.primeintellect.ai/traces/uploading.md) for `traces:read` / `traces:write`, workspace selection and the Python `prime-traces` SDK. Trace records and episode records are different formats: choose the format from the producer's records, not the `.jsonl` filename. Do not rename legacy prompt/completion/reward fields and assume they form a supported trace.

Review [ingestion/storage charges and automatic upload settings](https://docs.primeintellect.ai/traces/pricing-and-settings.md) before running jobs. Online SDK runs may upload automatically; writing a local file alone does not upload it. See [trace management and troubleshooting](https://docs.primeintellect.ai/traces/managing-traces.md) and the [trace upload API](https://docs.primeintellect.ai/api-reference/traces/ingestion/upload-traces.md). The [legacy storage retirement notice](https://docs.primeintellect.ai/traces/legacy-storage.md) is separate from the retired shared LoRA training service.


## Make your first inference request

Create a key with **Inference** permission. Set `PRIME_API_KEY` securely on the server. The OpenAI-compatible base URL is `https://api.pinference.ai/api/v1`, not the compute API host.

Find a current model ID with `prime inference models` or the [Models API](https://docs.primeintellect.ai/api-reference/inference-models.md). Check the [live catalog](https://app.primeintellect.ai/dashboard/inference?tab=models) for pricing, limits, capabilities and whether a model is Prime-hosted or routed to an external provider through the gateway.

The following request is billable. Replace `MODEL_ID_FROM_CATALOG` with an exact current model ID:

```bash
curl https://api.pinference.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $PRIME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"MODEL_ID_FROM_CATALOG","messages":[{"role":"user","content":"Hello"}]}'
```

For team billing add `-H "X-Prime-Team-ID: YOUR_TEAM_ID"` to each request. Without it, requests bill the API key owner's personal account. With the OpenAI Python SDK:

```python
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["PRIME_API_KEY"],
    base_url="https://api.pinference.ai/api/v1",
    # For team billing, add default_headers={"X-Prime-Team-ID": "YOUR_TEAM_ID"}.
)
response = client.chat.completions.create(
    model="MODEL_ID_FROM_CATALOG",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)
```

Check the live model's supported features before relying on tools, structured output, streaming or context limits. Evaluation models such as Clef use the Evaluations API rather than chat completions. For custom-model dedicated deployments, [contact the team](/contact) about capacity and deployment requirements.

### Inference reference

- [Quickstart, hosting types and team billing](https://docs.primeintellect.ai/inference/overview.md)
- [Models API](https://docs.primeintellect.ai/api-reference/inference-models.md)
- [Chat Completions API](https://docs.primeintellect.ai/api-reference/inference-chat-completions.md)
- [Evaluations API](https://docs.primeintellect.ai/api-reference/inference-evaluations.md)
- [Troubleshooting](https://docs.primeintellect.ai/inference/troubleshooting.md): check the selected billing account for insufficient-funds errors
- [Evaluate your model on an environment](https://docs.primeintellect.ai/tutorials-environments/evaluating.md)


## Provision GPU compute

1. Configure your account, billing and SSH key using the [CLI setup guide](https://docs.primeintellect.ai/cli-reference/introduction.md). Use `prime config set-ssh-key-path` for pod access.
2. Query current capacity and price before creating a pod:

```bash
prime availability list
prime availability list --gpu-type H100_80GB
```

3. Follow the [provisioning guide](https://docs.primeintellect.ai/cli-reference/provision-gpu.md) to choose provider, region, GPU count, image and storage. `prime pods create` provisions paid capacity; do not create it until the workload and budget are approved.
4. Inspect instances with `prime pods list`, and connect using `prime pods ssh <pod-id>`. Follow [pod lifecycle documentation](https://docs.primeintellect.ai/api-reference/managing-pods.md) to terminate resources when finished.
5. Put data you need to retain on [persistent storage](https://docs.primeintellect.ai/tutorials-storage/create-persistent-storage.md) and [attach it correctly](https://docs.primeintellect.ai/tutorials-storage/use-persistent-storage-with-instances.md). Review storage charges and retention separately from compute.

The compute API uses `https://api.primeintellect.ai`. Start with the [availability API](https://docs.primeintellect.ai/api-reference/check-gpu-availability.md), then [provisioning API](https://docs.primeintellect.ai/api-reference/provision-gpu.md). Availability read permission is required for capacity queries. Use provider identifiers, images and datacenter fields from the response rather than inventing them; account for disk and other resource charges as well as the GPU rate.

### Dedicated clusters

[Contact our team](/contact) for a reserved cluster. Read [Model Factory clusters](https://docs.primeintellect.ai/model-factory/clusters.md), [monitoring](https://docs.primeintellect.ai/model-factory/monitoring.md) and [Slurm orchestration](https://docs.primeintellect.ai/model-factory/slurm-orchestration.md) for the managed training/inference stack. Do not assume on-demand pod configuration applies to a dedicated cluster.

## Run code in a sandbox

Sandboxes are isolated microVMs booting OCI/Docker images. Authenticate with sandbox permissions, then follow this billable quickstart:

```bash
prime sandbox create python:3.11-slim --timeout-minutes 120
prime sandbox list
prime sandbox run <sandbox-id> "python --version"
prime sandbox delete <sandbox-id>
```

Replace `<sandbox-id>` with the returned ID. Delete the sandbox when done, including after a failed task. Set explicit lifetime/idle timeouts and resource limits. GPU sandbox access requires an explicit grant; do not assume a GPU flag makes it available.

A filesystem checkpoint is not the same as pause/resume: [checkpoints and forks](https://docs.primeintellect.ai/sandboxes/checkpoints.md) save filesystem state; [pause/resume](https://docs.primeintellect.ai/sandboxes/pause-resume.md) preserves memory and processes. Paused sandboxes still incur disk charges; images, volumes and checkpoints may retain storage charges after compute stops. Read the [current pricing and limits](https://docs.primeintellect.ai/sandboxes/overview.md) rather than relying on fixed prices in this summary.

### Sandbox reference

- [Lifecycle, resources and cleanup](https://docs.primeintellect.ai/sandboxes/lifecycle.md)
- [Commands and SSH](https://docs.primeintellect.ai/sandboxes/commands.md)
- [File transfer](https://docs.primeintellect.ai/sandboxes/files.md)
- [Secrets](https://docs.primeintellect.ai/sandboxes/secrets.md)
- [Outbound network isolation](https://docs.primeintellect.ai/sandboxes/network-isolation.md)
- [Private networks and public ports](https://docs.primeintellect.ai/sandboxes/networks.md)
- [Persistent volumes](https://docs.primeintellect.ai/sandboxes/volumes.md)
- [Custom images](https://docs.primeintellect.ai/sandboxes/images.md)
- [Tunnels](https://docs.primeintellect.ai/sandboxes/tunnel.md)
- [Python SDK: sync/async clients and error handling](https://docs.primeintellect.ai/sandboxes/sdk.md)
- [Runnable SDK examples](https://github.com/PrimeIntellect-ai/prime/tree/main/examples)


## Open-source tools

- [Prime CLI and Python SDKs](https://github.com/PrimeIntellect-ai/prime): platform clients and runnable examples.
- [Verifiers](https://github.com/PrimeIntellect-ai/verifiers): environments and evaluations; start with the [overview](https://docs.primeintellect.ai/verifiers/overview.md) and [v1 quickstart](https://docs.primeintellect.ai/verifiers/v1/getting_started.md).
- [prime-rl](https://github.com/PrimeIntellect-ai/prime-rl): self-hosted asynchronous RL; read [configuration](https://docs.primeintellect.ai/prime-rl/configuration.md), [training](https://docs.primeintellect.ai/prime-rl/training.md) and [scaling](https://docs.primeintellect.ai/prime-rl/scaling.md).
- [Renderers](https://github.com/PrimeIntellect-ai/renderers): model-specific message/token rendering for training workflows.

Use the documentation matching your installed version. Hosted services and self-hosted libraries have different access, deployment and billing requirements.

## Before an agent takes action

Confirm the user's workspace, region/data constraints, budget and intended resource lifetime before launching paid jobs. Never silently create resources, upload private datasets/traces, publish an endpoint or delete retained storage. Keep credentials out of logs and traces. Verify job status and saved outputs; an accepted API request is not proof that a job finished. On failure, inspect the documented error and current state before retrying a write. Clean up compute explicitly, and confirm separately whether persistent storage should remain.

## Updates, support and company information

- [Product changelog in Markdown](/changelog.md) or [on the website](/changelog): dated releases and product updates. Use live documentation for current behavior and pricing.
- [Prime CLI/SDK releases](https://github.com/PrimeIntellect-ai/prime/releases)
- [Verifiers releases](https://github.com/PrimeIntellect-ai/verifiers/releases)
- [prime-rl releases](https://github.com/PrimeIntellect-ai/prime-rl/releases)
- [Blog and research](/blog) and [case studies](/case-study)
- [Contact sales/support](/contact) or [email contact@primeintellect.ai](mailto:contact@primeintellect.ai) for beta access, dedicated deployments and higher limits
- [Security](/security), [Privacy Policy](/privacy-policy) and [Platform Terms](/terms-of-service)
- [Careers](/careers)