PrimeIntellect-ai/prime

Prime Intellect CLI & SDKs – Overview

What it is – A Python‑based command‑line tool and a set of SDKs that let you work with Prime Intellect’s cloud platform. The platform offers hosted model training, evaluation, GPU‑pod provisioning, sandboxed code execution, and a shared hub of verified environments. All interactions are done through the prime CLI or the lightweight SDK packages (prime‑sandboxes, prime‑evals, prime‑tunnel).

Key capabilities

• Lab workspaces – Local folders pre‑populated with config files, verifiers, and example RL/SFT configs, making it easy to develop and test against the cloud services.
• Hosted training – Submit reinforcement‑learning or supervised‑fine‑tuning jobs (prime train …) that run on dedicated GPU clusters. You can view logs, metrics, and checkpoints from the CLI.
• Environments hub – Browse, inspect, install, or push “verified” environments (e.g., gym‑style RL worlds) that are tightly integrated with the training and evaluation pipelines.
• Evaluations – Push evaluation results to the hub, list them, and inspect samples (prime eval …).
• GPU resource management – Query available GPU types and capacity, filter by model (e.g., H100_80GB).
• Pod management – Create, monitor, SSH into, and terminate compute pods that run your code in the cloud.
• Sandboxes – A tiny SDK (prime‑sandboxes) for running AI‑generated code safely in isolated containers.
• Team support – Switch between personal and team contexts, list teams, and manage shared resources.

How to get started

1. Install the modern Python package manager uv and then the CLI: uv tool install prime (or pip install prime).
2. Log in with prime login to obtain an API key.
3. Run prime lab setup to create a workspace with starter configs.
4. Use prime train models to see what training models are available, then launch a job with prime train <config>.toml.
5. Explore environments (prime env list), GPU availability (prime availability list), or spin up a pod (prime pods create).

Who it’s for – Researchers, ML engineers, or teams that want a managed service for training RL agents or fine‑tuning LLMs without handling their own GPU fleet. The CLI abstracts cloud provisioning, while the SDKs let you embed the same functionality in Python scripts.

Installation options

• Recommended: uv tool install prime (fast, isolates the tool).
• Standard pip install works as well.
• For sandbox‑only use, install the tiny prime-sandboxes package (~50 KB).

Development – The repo is a monorepo with four independently versioned packages. Developers can clone, run uv sync, and install the CLI in editable mode to contribute or run the test suite.

Links – Website, dashboard, API docs, and a Discord community are provided for further help.

---

All details are taken directly from the README; no additional claims are made.