OpenPipe

license

Apache-2.0(OSI, OpenPipe Inc.

source

public

methods

RL(GRPO)+SFT trajectory training

managed-tier

OpenPipe hosted training/serving SaaS on top of the OSS lib

core features withheld

no — ART FAQ commits to maintaining ART as a full-featured open source project and calls the hosted ART backend optional, for users who do not want to manage GPU infrastructure

OpenPipe's ART library is Apache-2.0 with the full source public, and nothing is withheld: no enterprise directory, no closed dependency, no license key. The hosted training and serving platform is a separate product sold alongside the open core, and a fully self-hostable LocalBackend means hosting does not gate any ART feature.

• https://art.openpipe.ai/fundamentals/art-backend recorded 2026-10-14

  Documents two interchangeable backend classes on the same client code: ServerlessBackend "train remotely on autoscaling GPUs" and LocalBackend "run your agent and training code on the same machine" — no capability described as exclusive to either.

• https://art.openpipe.ai/getting-started/faq recorded 2026-10-14

  ART FAQ states "We are committed to maintaining ART as a full-featured open source project" and "We've included an open-source ART backend". On the hosted option it says "We will also deploy an optional hosted ART backend for users who don't want to manage GPU infrastructure on their own." The FAQ names no paid OpenPipe plan and no feature paywall.

• https://github.com/OpenPipe/ART recorded 2026-09-09

  Repo metadata reads `"isPrivate":false`, `"isArchived":false`, `"visibilityLabel":"Public"`, and `"license":{"spdxId":"Apache-2.0","name":"Apache License 2.0"}`.

• https://raw.githubusercontent.com/OpenPipe/ART/main/LICENSE recorded 2026-10-14

  Full Apache License 2.0 text; appendix reads "Copyright 2025, OpenPipe, Inc." followed by an "Additional License Notice" stating the distribution "includes a small portion of code adapted from a third-party project licensed under the GNU Lesser General Public License v3.0."

• https://raw.githubusercontent.com/OpenPipe/ART/main/README.md recorded 2026-10-14

  "Train from anywhere. Run the ART client on your laptop and let the ART server kick off an ephemeral GPU-enabled environment, or run on a local GPU." Install is `pip install openpipe-art`.

Adoption is measured by PyPI downloads of openpipe-art, averaged over pypistats' overall series. GitHub stars corroborate the picture but do not set the reading, since a measured download figure is a better use signal than a star count.

• https://pypistats.org/packages/openpipe-art recorded 2026-08-09

  The page's "Overall" Plotly series for openpipe-art covers 2026-02-09 through 2026-08-09; summing its daily values over the last 30 recorded days (through 2026-08-07) gives 22,537 downloads.

ART is a focused agent-RL trainer that trains LoRA adapters with GRPO on top of vLLM and Unsloth, and its documentation describes no tensor, pipeline or expert parallelism. That leaves it well behind Megatron-LM, which trains models of hundreds of billions of parameters across thousands of GPUs.

• https://art.openpipe.ai/fundamentals/art-backend recorded 2026-09-28

  The ART backend page documents ServerlessBackend "train remotely on autoscaling GPUs" and LocalBackend "run your agent and training code on the same machine". Searched the body for tensor parallel, pipeline parallel, expert parallel, Megatron, multi-GPU, FSDP and DeepSpeed: none found.

• https://art.openpipe.ai/getting-started/faq recorded 2026-09-28

  The FAQ says ART "currently builds on GRPO, an RL algorithm that involves running many agents in parallel and then using the difference in the rewards they achieve as a stable training signal". Searched the body for tensor parallel, pipeline parallel, expert parallel, Megatron, MLPerf, multi-GPU, FSDP and DeepSpeed: none found.

• https://raw.githubusercontent.com/OpenPipe/ART/main/README.md recorded 2026-09-28

  The README describes ART as "an open-source RL framework that improves agent reliability by allowing LLMs to **learn from experience**" built on GRPO, where the server "runs the model's latest LoRA in vLLM" and trains it with GRPO, and says it "should work with most vLLM/HuggingFace-transformers compatible causal language models". Searched the body for tensor parallel, pipeline parallel, expert parallel, Megatron, MLPerf, multi-GPU, FSDP and DeepSpeed: none found.