NeMo-RL

license

Apache-2.0(OSI)

source

public

maintainer

NVIDIA

core-gated

ungated

Apache-2.0, an OSI license, with the full source public and no enterprise or paid directory anywhere in the repository. NVIDIA's separate NeMo Customizer microservice, part of NeMo Microservices, withholds nothing from the published code. The project is oriented to NVIDIA GPUs, but that is a hardware practicality rather than a license restriction.

• https://github.com/NVIDIA-NeMo/RL recorded 2026-08-09

  README opens "NeMo RL is an open-source post-training library under the NVIDIA NeMo Framework, designed to streamline and scale reinforcement learning methods for multimodal models (LLMs, VLMs etc.)" and its Bare-Metal Quick Start documents "git clone git@github.com:NVIDIA-NeMo/RL.git" as an alternative to the NGC container, so the thing you run builds from the published source. The file tree lists 30 top-level items - directories .agents/contributor-skills, .claude, .github, 3rdparty, docker, docs, examples, infra, nemo_rl, research, skills, tests, tools, plus files including LICENSE, README.md, CONTRIBUTING.md, pyproject.toml and uv.lock - with no enterprise, commercial or paid-edition directory. The repo's license badge reads "Apache-2.0 license" and GitHub's own license classifier records spdxId "Apache-2.0".

• https://raw.githubusercontent.com/NVIDIA-NeMo/RL/main/LICENSE recorded 2026-08-09

  The fetched body is the unmodified Apache License, Version 2.0 text end to end - Sections 1 through 9 plus the appendix template with its placeholder "Copyright [yyyy] [name of copyright owner]" line still unfilled - an OSI-approved grant of a perpetual, worldwide, royalty-free copyright and patent license with no restriction on who may use the software or at what scale, only the Section 4 attribution/change-notice/NOTICE obligations.

• https://developer.nvidia.com/nemo-customizer recorded 2026-08-09

  The URL redirects to docs.nvidia.com/nemo/microservices/latest/customizer/index.html, NVIDIA's documentation for "NeMo Customizer" under the separate NeMo Microservices platform; its own banner points to a distinct GitHub repo, "NVIDIA-NeMo/nemo-platform", not NVIDIA-NeMo/RL. A case-insensitive search of the saved 85,344-byte body for 'nemo-rl' or 'nemo rl' returns 0 matches for each. Nothing on the page describes a feature withheld from the nemo-rl repository; Customizer reads as a separately branded managed microservice rather than a paid tier of nemo-rl.

GitHub stars are the only usable signal here: 1,890 of them, displayed as 1.9k, which falls in the 1K-10K band. The "Used by" and contributor counts are not exposed on the page, and the nemo-rl package on PyPI is a non-functional placeholder at version 0.0.0 that directs users back to GitHub, so no download figure exists to band on. The level therefore rests on a weaker signal than most entries here.

• https://github.com/NVIDIA-NeMo/RL recorded 2026-08-09

  Sidebar records 1,890 stars (displayed '1.9k', up from the ~1.7k previously recorded), 501 forks, 45 watchers, and a commit count of 1,301. The 'Used by' and contributor-count sidebar sections are unloaded skeleton placeholders in the static page (populated by client-side JavaScript), so no usable count is available from either.

• https://pypi.org/project/nemo-rl/ recorded 2026-08-09

  The project page for the 'nemo-rl' PyPI listing reads: "WARNING: This package is not functional and a placeholder from NVIDIA. Please refer to https://github.com/NVIDIA/NeMo-RL for instructions on installing and using NeMo-RL." It is version 0.0.0, released Jun 3, 2025, a 6.0 kB source distribution with no functional code, so PyPI download counts cannot serve as an adoption signal for this product.

Frontier-tier scale and parallelism for a fine-tuning library: DTensor backends covering FSDP2 and tensor, sequence, pipeline and context parallelism through NeMo AutoModel, and Megatron Core with six-dimensional parallelism through NeMo Megatron Bridge. Method coverage is strong too -- GRPO, GSPO, DAPO and GDPO, SFT and DPO with LoRA, on-policy distillation and reward modeling -- alongside an SGLang backend, speculative decoding and long-context training. Capability is high even though adoption is still small.

• https://github.com/NVIDIA-NeMo/RL recorded 2026-08-09

  README's Features list includes "Learning Algorithms - GRPO/GSPO/DAPO, SFT(with LoRA), DPO, and On-policy distillation", a Table of Contents with RM (reward modeling) Single-Node/Multi-node sections, "Advanced Parallelism with DTensor... PyTorch FSDP2, TP, CP, and SP... (through NeMo AutoModel)", "Larger Model Support with Longer Sequences - Performant parallelisms with Megatron Core (TP/PP/CP/SP/EP/FSDP) (through NeMo Megatron Bridge)", plus SGLang Inference, Speculative Decoding and GDPO Support news entries - all matching the recorded feature matrix. No 'PPO' string and no 'H100' or 'B200' string appear anywhere in the README; only GB200 container support is named.