AReno
AReno is a self-contained toolkit for running reinforcement-learning post-training of language models, PPO, GRPO and SFT or DPO-style training, serving and agentic RL, on a single node without a training cluster.
AReno is a local LLM post-training toolkit developed by engineers from Ant Group's ASystem team, aimed at going from a base checkpoint to a trained, served model on one machine without standing up a cluster or wiring together separate training, inference and kernel components. Post-training methods, such as PPO, GRPO or SFT and DPO-style training, are selected through an --algo flag or by using the same Trainer class directly from Python.
Its self-contained train-and-serve stack installs only the native backend the host needs, CUDA on Linux or MLX on Apple Silicon, and selects between them automatically without silently falling back from one to the other. Agentic RL support runs a user's agent function against AReno's own local OpenAI-compatible proxy, then trains from the resulting trajectories, tokens, log probabilities, rewards and loss masks. Native LoRA support trains TP-aware adapters for Qwen3, Qwen3-MoE and Bailing-MoE V3 models and saves them as standard PEFT artifacts that reload for further training or serving.