AReaL-MinT¶
AReaL-MinT is an open-source, self-hosted MinT training runtime built on AReaL (FSDP2 training) and SGLang (inference). It exposes the same HTTP API as MinT, so the same SDK script runs unchanged against a local deployment.
Under the hood, a FastAPI control plane dispatches each request to its backend: training to AReaL's FSDP2 engine, inference to SGLang with LoRA hot-reload. Every write returns a future the client polls.
MinT exposes a Tinker-compatible client surface β see the Tinker-compatible API reference for the 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¶
For prerequisites, install steps, and the architecture overview, see the AReaL-MinT repository.