AReaL-MinT

AReaL-MinT is an open-source Mint Recursive training runtime you can deploy yourself. It uses AReaL's FSDP2 engine for training and SGLang for inference. Its HTTP API matches Mint Recursive's, so your existing SDK scripts run unchanged on a local deployment.

A FastAPI control plane routes training requests to AReaL's FSDP2 engine and inference requests to SGLang, which supports LoRA hot-reload. Each write returns a future that the client polls for the result.

Mint Recursive's client interface is compatible with Tinker. See the Tinker-compatible API reference for shared types and clients.

Algorithms

Algorithm	Status	Notes
SFT	Available	LoRA SFT, checkpoint save/load, resume, sampler handoff
GRPO	Available	Group Relative Policy Optimization; group-relative advantages without a value critic

Supported models

Family	Size	GPU layout
Qwen3.6	27B (Dense)	4 FSDP actor + 4 SGLang rollout
Qwen3.6	35B-A3B (MoE)	4 FSDP actor + 4 SGLang rollout

Quick start

See the AReaL-MinT repository for environment requirements, installation steps, and runtime architecture.