AI Potluck
Model components / Fine-tuning code

verl

ByteDance Seed / Volcano Engine

HybridFlow flexible and efficient RL post-training framework developed by ByteDance Seed MLSys and HKU, purpose-built for RLHF and reasoning RL on math and code tasks. Picked when teams need state-of-the-art GRPO/PPO at frontier scale; verl shipped early support for DeepSeek-R1-style reasoning RL recipes and has become a reference implementation in the post-DeepSeek wave. 21.5K GitHub stars, 618 contributors, ~420 commits in 90 days, one of the fastest-growing repos in the category.

verl (Volcano Engine RL / HybridFlow), v0.8.0 released June 1 2026. Initiated by ByteDance Seed, now community-maintained. RL post-training library (takes existing weights and adapts via RLHF/RL). Confirmed live on GitHub June 2026.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public
governance
community(ByteDance-Seed origin)
core-gated
ungated

Fully OSI-licensed; full source public, no managed-tier gating of the library itself.

Adoption

4 high confidence
4.0

~96.7k PyPI downloads last month (primary). De facto standard high-throughput RL post-training lib; used to train Seed-Thinking-v1.5 and Doubao, and adopted/contributed by Qwen team, Moonshot/Kimi, NVIDIA research, Microsoft Research, Amazon, LinkedIn, Xiaomi, Baidu and many labs. Level 4 reflects broad and growing developer-tool usage at the high end of the 100K-1M band trending toward heavier use.

Capability

5 high confidence
5.0

Frontier-tier within finetuning_code: broadest modern RL-algorithm coverage and demonstrated use training production frontier reasoning models. Calibrated alongside Megatron-LM (C5) as a scale/method definer for this category.

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