Model components / Fine-tuning code

OpenPipe

OpenPipe ships the open source Agent Reinforcement Trainer (ART, Apache-2.0, ~9.7K stars) alongside a closed managed prompt-logging and fine-tuning platform, an open-core stack for SFT and RL post-training of agents. Acquired by CoreWeave in September 2025 to anchor their RL-for-agents capability alongside Weights & Biases; ART continues active development as the open library, with the platform serving as the commercial wrapper.

Scored on ART (Agent Reinforcement Trainer), OpenPipe's open source fine-tuning/RL library (GRPO). v0.5.17 (Mar 13, 2026), confirmed live on GitHub June 2026. OpenPipe also runs a managed training/hosting platform on top of the OSS library.

Openness

4 high confidence

4.0

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-gated: gated

Core ART library is OSI-licensed (Apache-2.0) and fully public; OpenPipe's revenue product is a managed training/hosting platform layered on top, so open_core.

https://github.com/OpenPipe/ART/blob/main/LICENSE recorded 2026-06-04
Apache-2.0 license text (OpenPipe, Inc. 2025; LGPL-3.0 notice for vendored portions)
https://github.com/OpenPipe/ART recorded 2026-06-04
ART = open source GRPO RL trainer; v0.5.17 release; active dev

Adoption

2 medium confidence

2.0

openpipe-art PyPI ~47.7k downloads last month; 9.9k GitHub stars (corroborating only). Specialist RL-fine-tuning audience; 10K-100K monthly-download band.

https://pypistats.org/packages/openpipe-art recorded 2026-06-04
~47,684 downloads last month for openpipe-art
https://github.com/OpenPipe/ART recorded 2026-06-04
9.9k stars (secondary corroboration)

Capability

3 medium confidence

3.0

Capable, focused agent-RL trainer; mid-tier vs full training stacks (no TP/PP/EP at Megatron scale). Feature-matrix basis per recipe; no MLPerf-Training submission.

https://github.com/OpenPipe/ART recorded 2026-06-04
GRPO RL trainer, vLLM/Unsloth integration, Qwen/Llama/GPT-OSS support

Unchanged since 2026-07-30 (last edited, not re-checked)