What is AReno?

AReno is a compact toolkit for training and serving large language models using reinforcement learning on a single machine. Created by engineers at Ant Group's inclusionAI team, it bundles everything needed for post-training — multiple RL algorithms, an agent loop, efficient adapter methods, and a model server — into one package that runs on NVIDIA GPUs, Huawei NPUs, or Apple Silicon. You install it via a script, Docker, or pip, then use a command-line interface or Python SDK to run training jobs. A web dashboard can track experiments locally or on Modal's cloud platform. The code is Apache-2.0 licensed and under active development.

Should you self-host AReno?

AReno scores 49/100 on selfhost.directory (how scoring works) and ranks #53 of 125 in AI. It has 381 GitHub stars.

It shipped 10 stable releases in the past year. The latest, v0.10.0, came out on 23 October 2026. New versions land roughly every 2 weeks.

Setup is advanced: plan for several services and some ongoing maintenance. An official Docker image is available. It is Apache-2.0-licensed and written mostly in Python.

Its closest alternative on selfhost.directory is ComfyUI (93/100, 143 701 stars). It scores higher overall, so compare the two before you pick. See AReno vs ComfyUI side by side.

Features

Agentic RL
Auto Memory Tuning
Extensible plugin system
JEV Decision Heads
Multiple Hardware Backends
Native LoRA
OpenAI-Compatible Serving
QLoRA
RL Post-Training Algorithms

FAQ

Should examples be copied from Cookbook or Reference?

Start from Cookbook for runnable recipes. Use Reference when you already know which command, SDK type, or API contract you need.

Where should I start after installation?

Run the Quickstart guide, then choose the RLVR or agentic rollout path that matches your task.

Is FlashAttention required?

FlashAttention is optional unless the selected attention backend requires it. Use --attn-backend native when debugging unsupported FlashAttention setups.