OpenPipe

OpenPipe builds tools for adapting language models to specific applications, from capturing prompts and completions for fine-tuning to training multi-turn agents through reinforcement learning. Founded in 2023 by Kyle Corbitt and David Corbitt, it participated in Y Combinator’s Summer 2023 batch and was acquired by CoreWeave in 2025.[2][8][11]

Before its work on agent reinforcement learning, OpenPipe’s initial product addressed the cost and latency of existing LLM applications by using their recorded interactions to train specialized models.[2]

In their founder launch account, David described an app that searched Reddit and classified results along user-specific dimensions; the volume of GPT-3.5 classification made each search cost multiple dollars. Kyle had built and sold a startup involving official-document translation. He recalled difficulty getting GPT-3.5 to follow translation requirements reliably, while GPT-4 was too slow for a good user experience.[2]

OpenPipe’s SDK captured an existing provider’s prompt-completion pairs in the background. The launch workflow called for prototyping with GPT-3.5 or GPT-4, collecting a few hundred to a few thousand completions, and starting a training job in OpenPipe’s UI. Developers could download the resulting model for self-hosting or host it with OpenPipe. The founders presented the switch as an SDK URL change that retained the application’s prompts and code.[2]

OpenPipe also experimented with base models for fine-tuning. In a December 18, 2023 account of Mistral 7B Fine-Tune Optimized, Kyle reported testing four models formed by merging candidates, then fine-tuning each on three datasets.[6]

By March 26, 2024, OpenPipe had closed a $6.7 million seed round led by Costanoa Ventures, with participation from Y Combinator and angels including Logan Kilpatrick, Alex Graveley, and Tom Preston-Werner. At that time, its collaborators included Langfuse for LLM debugging and Athena AI for data collection; OctoAIML and Fireworks AI helped serve its fine-tuned models.[5]

OpenPipe’s customer examples included translating and extracting information from government regulations, categorizing and filtering YouTube videos using transcripts, and routing inbound customer-service requests. In the March 2024 financing report, SiliconANGLE relayed the company’s estimate that its smaller models had saved customers over $3 million in inference costs since the platform’s September launch.[5]

OpenPipe initially distilled expensive GPT-4 workflows into smaller, cheaper models, then pivoted from supervised fine-tuning to reinforcement learning for agents as frontier-model prices fell, according to Latent Space’s October 16, 2025 account.[1]

ART (Agent Reinforcement Trainer) is OpenPipe’s open-source framework for multi-turn LLM-agent reinforcement learning using GRPO, or Group Relative Policy Optimization. Its training server is a modular service separate from the application’s client code. The ART documentation describes running the client on a laptop while training executes in an ephemeral GPU environment or on a local GPU, with observability integrations for W&B, Langfuse, and OpenPipe and an autoscaling GPU integration through W&B Training.[8]

OpenPipe introduced RULER in a July 11, 2025 post by Kyle Corbitt. The method uses LLM judges to rank agent behaviors relatively rather than absolutely, according to Latent Space’s account. RULER-generated scores are used directly as rewards in a GRPO training step.[7][1]

Kyle Corbitt announced that OpenPipe was joining CoreWeave on September 3, 2025. GeekWire’s acquisition report said OpenPipe’s employees would join CoreWeave and reported a count of ten employees according to LinkedIn.[3][4]

The acquisition was completed: CoreWeave’s Form 10-Q for the quarter ended September 30, 2025 records the issuance of 272,169 Class A common shares to former OpenPipe common shareholders on September 5, 2025, in connection with the acquisition’s closing. That issuance does not establish the total acquisition consideration.[11]

OpenPipe subsequently stated on its official website that its core functionality, including model distillation, had been successfully migrated to the CoreWeave platform following the acquisition.[10]