# PrimeIntellect-ai/prime-rl

Agentic RL Training at Scale

Repository: https://github.com/PrimeIntellect-ai/prime-rl
Canonical: https://ross.abutalabs.com/products/prime-rl
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T19:54:30+00:00

## Health v2 (maintenance only)
Score: 87/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 99, longevity 40
- inputs: {"age_days": 561, "days_push": 7, "days_rel": 8, "gap_med": 22.5, "n_releases_24m": 9}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1975, forks 416 (observed 2026-08-28T04:06:00.966973+00:00)

## What it is
prime-rl is a Python framework for large-scale, fully asynchronous reinforcement learning training of language models, built on FSDP2 for training and vLLM for inference. It supports scaling to 1000+ GPUs, integrates with the Prime Intellect Environments Hub for agentic RL environments, and covers end-to-end post-training including SFT, RL, and evals.

## Use cases
- train LLMs with reinforcement learning at scale
- run agentic RL post-training on large MoE models
- fine-tune language models with SFT and RL pipelines
- deploy multi-node RL training jobs on Slurm or Kubernetes
- evaluate and post-train models on agentic environments like SWE
- train vision-language models with RL

## When to choose
- you need to scale RL training to hundreds or thousands of GPUs
- you want asynchronous, high-throughput agentic RL training
- you need integrated SFT, RL, and evaluation in one framework
- you want native integration with verifiers environments and the Environments Hub
- you need optimized support for large MoE models with expert and context parallelism

## When to avoid
- you only need simple single-GPU fine-tuning with minimal setup
- you need a lightweight RL library for small models or quick experiments
- you are not working with language models or agentic RL tasks
- you lack access to multi-GPU or multi-node infrastructure

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, reinforcement-learning, machine-learning, gpu-computing, benchmarking
- domain: reinforcement-learning, large-language-models, machine-learning, gpu-computing, artificial-intelligence
- platform: python, cloud
- tags: reinforcement-learning, rlhf, post-training, sft, vllm, fsdp2, slurm, agentic-training, moe, distributed-training, verifiers, environments-hub, multimodal, fp8, evals, gpu, linux, docker, kubernetes

## Member repositories
- PrimeIntellect-ai/prime-rl (main) score 87

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:00.966973+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:04:51.652278+00:00, confidence not recorded.
  - readme: https://github.com/PrimeIntellect-ai/prime-rl (fetched 2026-08-28T04:06:00.966973+00:00, sha 71f2fe33a853)
- Data as of 2026-08-30T08:39:29.467469+00:00.