prime-rl

license

Apache-2.0(OSI)

source

public(PrimeIntellect-ai/prime-rl)

core features withheld

no — the tree at cd26753 carries one root Apache-2.0 LICENSE and no ee/, enterprise/, commercial/ or proprietary/ path, and its five submodules are public repositories. Prime Intellect sells compute, hosted training and inference as separate platform products

prime-rl is Apache-2.0 licensed, and its public repository holds the trainer, orchestrator and inference server, which install and run from source with no account. Prime Intellect sells compute, hosted training and inference as separate products, and the framework's links to them, run reporting and hosted evaluation, are optional.

• https://raw.githubusercontent.com/PrimeIntellect-ai/prime-rl/main/docs/eval.md recorded 2026-10-19

  The eval guide says "By default the model is served by Prime Inference" and shows how "To evaluate a model you serve yourself" with a local vLLM server; the file monitor is on by default and --monitors.prime, which streams to the Prime platform, is opt-in.

• https://raw.githubusercontent.com/PrimeIntellect-ai/prime-rl/main/LICENSE recorded 2026-10-19

  The LICENSE is the stock Apache License 2.0 text, with the unfilled appendix line "Copyright [yyyy] [name of copyright owner]", followed by a "Third-party code attributions" note that some code is derived from Meta Platforms under a BSD-style license; no use restriction is added.

• https://raw.githubusercontent.com/PrimeIntellect-ai/prime-rl/main/README.md recorded 2026-10-19

  The README installs prime-rl by cloning the repository and running uv sync, validates SFT, inference and the full RL stack locally, closes "This project is licensed under the Apache 2.0 license", and recommends "our compute platform" for renting GPUs, a separate offering.

• https://www.primeintellect.ai/ recorded 2026-10-19

  Prime Intellect's homepage presents Lab ("Hosted Training", "Managed training workflows"), Compute and Inference as its platform products, and names the open-source Verifiers library as what its RL environments are built on.

No package registry carries prime-rl: it installs from a git checkout, and its release images go to GitHub's container registry, which publishes no pull count. GitHub stars are the only available signal, and a star is not a use, so confidence is low.

• https://raw.githubusercontent.com/PrimeIntellect-ai/prime-rl/main/.github/workflows/image.yml recorded 2026-10-19

  The image workflow sets REGISTRY: ghcr.io and IMAGE_NAME: primeintellect-ai/prime-rl, so release images are pushed to GitHub's container registry rather than Docker Hub.

• https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PrimeIntellect-ai%2Fprime-rl recorded 2026-10-19

  ecosyste.ms's copy of the GitHub repository metadata for PrimeIntellect-ai/prime-rl: stargazers_count 2166, archived false, license "apache-2.0".

prime-rl is built for asynchronous RL at frontier scale, training mixture-of-experts models on FSDP2 with expert and context parallelism and vLLM inference, with SFT, distillation and evaluation in the same stack. It sits a step below verl, which documents a wider algorithm catalog, including PPO with a critic, and more than one training backend.

• https://raw.githubusercontent.com/PrimeIntellect-ai/prime-rl/main/docs/algorithms.md recorded 2026-10-19

  The algorithms page lists grpo, max_rl, rae, hierarchical_grpo, opd, sft, opsd and echo, with an IPO loss and async off-policy training, and says "prime-rl hosts exactly one model: the trainable policy", so there is no critic model.

• https://raw.githubusercontent.com/PrimeIntellect-ai/prime-rl/main/docs/scaling.md recorded 2026-10-19

  The scaling guide covers taking prime-rl "from a single GPU to a 1000-GPU cluster" with FSDP, expert parallelism, context parallelism and multi-node Slurm launches.

• https://raw.githubusercontent.com/PrimeIntellect-ai/prime-rl/main/README.md recorded 2026-10-19

  The README describes "Fully asynchronous RL for high-throughput agentic training at scale", "built to train 1T+ MoE models on 1000+ GPUs with FSDP2 for training and vLLM for inference, with FP8 inference, PD disaggregation, EP and CP parallelism", "End-to-end post-training: SFT, RL training, and evals", Slurm deployment and multimodal Qwen3.5 support.