CoreWeave Serverless RL

What it does

Serverless RL takes a base language model, a reward function that scores an agent's outputs, and batches of agent trajectories, then produces a low-rank adapter (LoRA) trained so higher-scoring behavior becomes more likely. The adapter is stored as an artifact in a Weights & Biases account, and every checkpoint is hosted on Serverless Inference so it can be called like any other model.

Developers run it either through the Agent Reinforcement Trainer (ART) open-source framework, which manages rollouts, rewards and checkpoints, or by calling the Serverless Training API directly to create models, generate rollouts, log scored trajectories and start training jobs. RULER is a separate option that generates relative rewards with an LLM-as-judge ranking trajectories, without labeled data or a handwritten reward function. Compute scales with training and returns to zero when work finishes, so no training cluster must be provisioned.

This suits ML engineers and developers building multi-turn agents who want managed RL training without operating GPU infrastructure. Access requires a Forge account and an API key created under User settings. Limitations: Serverless RL is in public preview, pricing and usage limits must be checked in the separate usage documentation, and users must write or choose a reward function, since training optimizes whatever it rewards.

What it can do

• Trains LoRA adapters with reinforcement learning from batches of scored agent trajectories
• Accepts a user-defined reward function that defines the training task
• Hosts every trained checkpoint on Serverless Inference through a wandb-artifact model endpoint
• Offers the Agent Reinforcement Trainer open-source framework for rollouts, rewards and checkpoints
• Provides RULER, an LLM-as-judge method that ranks trajectory groups to generate relative rewards without labeled data
• Exposes the Serverless Training API for creating models, logging trajectories and polling training jobs

Works with

• OpenPipe ART framework
• Weights & Biases projects and artifacts
• OpenAI-compatible chat completions endpoint

Example workflows

Train an agent with ART

Illustrative example. Input: a base model, a reward function scoring agent outputs, and a Weights & Biases project. Steps: install ART, follow the quickstart, run rollouts and scoring, and let ART submit training. Output: a LoRA adapter stored as a Weights & Biases artifact, hosted on Serverless Inference at a wandb-artifact endpoint.

Call the Serverless Training API directly

Illustrative example. Input: an API key, a model created with POST /v1/preview/models, and scored trajectories. Steps: generate rollouts via chat completions, log them with POST /v1/preview/models/{model_id}/log, start a training job with POST /v1/preview/training-jobs, and poll its status. Output: a trained checkpoint addressable through the model endpoint for inference.

Before you choose

• Serverless RL is in public preview as stated on the docs page.
• Training optimizes whatever the reward function rewards, so the function defines the task.
• Delete checkpoints you no longer need to save storage, per the docs.

Listed under

• Coding
• Sandboxes and infrastructure
• AI developers
• Developers
• Fine-tuning
• GPU infrastructure
• GPU training
• LLM agents

Get started

Create an account, then an API key under User settings in Forge; start with the ART quickstart or the API.

1. Create a Forge account at id.coreweave.com/signup and generate an API key under User settings in Forge, copying it when shown.
2. Create a Weights & Biases project to record training metrics and store the trained adapter.
3. Write a reward function that scores agent outputs, then install ART following its quickstart or call the Serverless Training API directly.

Serverless RL is in public preview. Confirm training, inference and artifact-storage charges and current usage limits before running jobs. Forge software subscription prices are not RL training rates.

We couldn’t find published prices for CoreWeave Serverless RL. See their pricing page ↗

Sources for this overview

Product pages, documentation, and pricing checked 29 Oct 2026. Each source supports part of the overview; availability and prices can change.

• CoreWeave Serverless RL | Train and Improve AI Agents ↗www.coreweave.com
• CoreWeave Forge Pricing ↗www.coreweave.com
• Use Serverless RL - CoreWeave Docs ↗docs.coreweave.com

Company website

coreweave.com: product, pricing and policy pages

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 29 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.