Speed and Flexibility: Introducing the WorldJen SDK and CLI

Mar 26, 2026
Worldjen

TL;DR

New release: WorldJen SDK and CLI on PyPI.
Read more about WorldJen: https://moonmath.ai/posts/introducing-worldjen/
The CLI makes account and runner operations scriptable for fast, agentic workflows.
The SDK enables direct programmatic runs from Python pipelines with strong flexibility and concurrency.

WorldJen provides high-performance benchmarks for video and world models. It empowers researchers and builders with a unified platform for accurate, fast, and clear evaluations across dimensions like subject consistency, motion smoothness, physical mechanics, and more.

We are excited to announce a major milestone in the evolution of WorldJen: the release of the WorldJen SDK and CLI. These tools fundamentally change how developers and researchers interact with the platform, accelerating development workflows and providing deeper control over experimental runs.

The WorldJen CLI: Programmatic Control from the Terminal

The WorldJen CLI brings full platform control to the terminal, enabling runs and benchmarks to be created, managed, and automated from the command line. This allows faster iteration on models and configurations, turning the account into a programmable system that fits existing workflows.^[1]

Infrastructure setup has also been simplified. The WorldJen Runner can now be registered and installed directly from the CLI, removing the need for external scripts and making it easier to connect custom compute to the platform.

The WorldJen SDK: Flexible, High-Performance Execution

The WorldJen SDK is designed for maximum control and performance. Runs can be executed using a thin Python wrapper around existing pipelines and configurations, enabling highly flexible workflows.^[2]

The SDK also supports concurrent execution across multi-GPU environments, whether the pipeline is using multiGPU generation or running concurrent pipelines, reducing the time required for large-scale evaluations.

A Look Ahead

We are already working on bringing this same level of speed and configuration directly to the WorldJen Runner:

Multi-runner single machine support
Deploying custom pipelines directly to a runner from CLI
Additional integrations for runners: GitHub repos, Weights and Biases, and more

Getting Started

The WorldJen SDK and CLI are now available on PyPI.

Watch the demo video on how to use WorldJen SDK and CLI below.

Before being able to use either the CLI or the SDK, you will need to create an API key in your account settings via the web UI; support for adding browser login via the CLI is already underway.

Once you have your API key, set it as the WORLDJEN_API_KEY environment variable either for the session or for the command:

# session
export WORLDJEN_API_KEY="wjk_my/api/key"

# single command
WORLDJEN_API_KEY="wjk_my/api/key" worldjen runs list

Then download the worldjen Python package from PyPI. If you are planning on installing the runner, include the optional runner dependencies as well:

uv pip install worldjen

# For installing the runner, you need the runner dependencies as well
uv pip install "worldjen[runner]"

Now you can use any command in the CLI and the SDK will pick up the exported environment variable. Check out our full demo above, or if you want a hands-on tutorial with LTX2, follow our guide to walk through the demo yourself.^[3]

Read more, automate your workflows, and accelerate your generative AI development!

TL;DR

New release: WorldJen SDK and CLI on PyPI.
The CLI makes account and runner operations scriptable for fast, agentic workflows.
The SDK enables direct programmatic runs from Python pipelines with strong flexibility and concurrency.
Deep dive into WorldJen: https://moonmath.ai/posts/introducing-worldjen/

Quick Example: Creating a Run via CLI

# Create a new run on a specific runner and model
worldjen runs create \
  --name "cli-experiment-v1" \
  --dimensions "dynamic_degree,motion_smoothness" \
  --runner-id "your-runner-id" \
  --model-id "your-model-id"

Quick Example: Running an LTX-Video Benchmark

import worldjen
from worldjen import Dimensions, run
from diffusers import LTXPipeline
import torch

# Configure your API key (or via WORLDJEN_API_KEY="wjk_mykey" env var)
worldjen.config(api_key="your-api-key")

# Load your pipeline
pipe = LTXPipeline.from_pretrained(
    "Lightricks/LTX-Video",
    torch_dtype=torch.bfloat16
).to("cuda:1")

# Execute a run directly from your code
result = run(
    pipe,
    dimensions=[Dimensions.DYNAMIC_DEGREE],
    run_name="sdk-example-ltx",
    model_id="Lightricks/LTX-Video",
    # optionally wait for the run to finish if you need the results afterwards
    wait_for_results=True,
    # Add any keyword arguments that your pipeline should receive aside from the prompt
    num_inference_steps=25
)

print(f"Run started: {result.run_id}")

This tutorial can be run on almost any machine. It works on macOS (including Apple Silicon), Linux, and systems with NVIDIA GPUs. Performance on Apple Silicon will be slower, and CPU-only runs may be very slow, but the tutorial is fully supported.