Back to Gap Map Model components / Fine-tuning code

verl

open source / Overall score: 3.5

verl (Volcano Engine Reinforcement Learning) is a post-training library for large language models, built on the HybridFlow hybrid-controller architecture. It targets RLHF and reasoning RL on math and code tasks, implements PPO, GRPO, RLOO, GSPO, and DAPO, and runs on FSDP or Megatron-LM with vLLM and SGLang rollout. ByteDance Seed started the project; the verl community maintains it.

The repository moved from volcengine/verl to the verl-project organization; the original path is still used for evidence.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI)
source: public
core features withheld: no

verl is released under the unmodified Apache-2.0 license, and its training and rollout code is in a public repository. Nothing the README documents, including the algorithms and training backends, sits behind a paid or enterprise edition.

https://github.com/volcengine/verl recorded 2026-08-09
Rendered repo page for the URL on file. The request 301-redirected to https://github.com/verl-project/verl (final_url), and the embedded repo metadata reports "license":{"spdxId":"Apache-2.0","name":"Apache License 2.0"}. The redirect matches the README's January 2026 org migration note ("verl has been migrated to the verl-project"); repository content is otherwise the same.
https://pypi.org/pypi/verl/json recorded 2026-09-28
PyPI metadata for verl reports "license": "Apache-2.0" and author "Bytedance - Seed - MLSys".
https://raw.githubusercontent.com/volcengine/verl/main/LICENSE recorded 2026-09-28
The LICENSE file is the Apache License text; its appendix still carries the unfilled placeholder "Copyright [yyyy] [name of copyright owner]" followed by "Licensed under the Apache License, Version 2.0 (the "License");" with no added use restriction.
https://raw.githubusercontent.com/volcengine/verl/main/pyproject.toml recorded 2026-09-28
The pyproject describes the package as "verl: Volcano Engine Reinforcement Learning for LLM" and its license field reads "Apache-2.0"; there is no separate closed distribution.
https://raw.githubusercontent.com/volcengine/verl/main/README.md recorded 2026-09-28
The README says "verl is the open-source version of" the HybridFlow paper, describes it as "a flexible, efficient and production-ready RL training library for large language models (LLMs)", and lists the full training and rollout feature set with no paid, enterprise or hosted tier; searched for enterprise, commercial, paid and premium: none.

Adoption

2 high confidence

2.0

Adoption is measured by PyPI downloads of verl. It is a reinforcement-learning training framework, so its audience is model-training teams rather than a mass developer base.

https://pypistats.org/api/packages/verl/recent recorded 2026-08-12
82,590 downloads in the trailing 30 days for verl

Capability

5 high confidence

5.0

verl supports a wide range of reinforcement-learning algorithms, including PPO, GRPO, RLOO, ReMax, REINFORCE++, GSPO, PRIME, DAPO and DrGRPO alongside supervised fine-tuning. It is built on the HybridFlow hybrid controller with FSDP and Megatron training backends and vLLM or SGLang rollout, scales to 671B-parameter models, and was used to train ByteDance's Seed-Thinking-v1.5, which puts it level with Megatron-LM.

https://raw.githubusercontent.com/volcengine/verl/main/README.md recorded 2026-09-28
The README lists "Scales up to 671B models and hundreds of GPUs" with FSDP, FSDP2 and Megatron-LM training, vLLM and SGLang rollout, and "Reinforcement learning with [PPO](examples/ppo_trainer/)" alongside GRPO, GSPO, ReMax, REINFORCE++, RLOO, PRIME, DAPO and DrGRPO plus supervised fine-tuning; it also reports that "Seed-Thinking-v1.5" was trained with verl.

Verified 2026-09-24