MinT: an RL training API for researchers and developers

MinT lets you focus on what matters in LLM post-training — your data, your loss functions, and your RL environments — while we handle the heavy lifting of distributed training.

You write a simple loop that runs on your CPU-only machine, including the data or environment and the loss function. We figure out how to make the training work on a bunch of GPUs, doing the exact computation you specified, efficiently. To change the model you're working with, you only need to change a single string in your code.

MinT gives you full control over the training loop and all the algorithmic details. It's not a magic black box that makes fine-tuning "easy". It's a clean abstraction that shields you from the complexity of distributed training while preserving your control.

Here's how the division of responsibilities works in practice:

Features

• Fine-tune open-weight models from 0.6B to 235B+ parameters
• Support for vision-language models like Qwen3-VL
• LoRA fine-tuning for efficient training of large models
• Distributed data collection and rolling training
• Weight management, model publishing, and download
• Online evaluation on standard CPU clusters

Core API Functions

• forward_backward: Compute and accumulate gradients
• optim_step: Update model parameters using accumulated gradients
• sample: Generate outputs from trained models
• save_state / load_state: Persist and restore weight and optimizer state
• save_weights_and_get_sampling_client: Publish weights for inference

Tinker Compatible

MinT is API-compatible with ThinkingMachines Tinker. If you have existing Tinker code, you can migrate to MinT with minimal changes.

# Option 1: Use MinT directly
import mint
client = mint.ServiceClient()

# Option 2: Use MinT as a Tinker drop-in
import mint as tinker
client = tinker.ServiceClient()

For detailed migration instructions and current compatibility status, see Tinker Compatibility.