Foundation models like Llama, Qwen, and Stable Diffusion are powerful but generic. Post-training is how you make them yours — teaching a model your data, your domain, your task.

That usually means training scripts, hyperparameter tuning, and GPU infrastructure. Gradients replaces all of it with a single API call. Choose a model and a dataset — Gradients automatically optimizes the training configuration for your problem and delivers a production-ready model.

pip install gradientsio

Sign up at gradients.io, grab an API key, and:

import os
from gradientsio import GradientsClient, TaskType

os.environ["GRADIENTS_API_KEY"] = "your-api-key"

result = GradientsClient().train(
    model="Qwen/Qwen2.5-3B",
    task_type=TaskType.INSTRUCT,
    hours=2,
    dataset="yahma/alpaca-cleaned",
    field_instruction="instruction",
    field_input="input",
    field_output="output",
).wait()

print(result.trained_model_repository)

Five training modes: INSTRUCT, CHAT, DPO, GRPO, IMAGE.

Model size	Hourly rate
Up to 1B parameters	$10 / hr	40B+ parameters	$50 / hr
Up to 7B parameters	$15 / hr	Image models	$5 / hr
Up to 40B parameters	$25 / hr

Getting Started	Install, create an account, train your first model, and test the result.
Task Types	Instruct, Chat, DPO, GRPO, and Image — when to use each and how.
Datasets	How to prepare your data, with good and bad examples for each type.
Configuration	Every parameter, error handling, polling, and environment variables.
Scheduler	Multi-iteration training for datasets larger than 300k rows.
Inference	Sample from local or cloud vLLM servers.
Deployment	Serve base models and LoRA adapters locally, on RunPod, Lium, Targon, or Basilica.
Account	Billing, balance, and pricing.

See the training comparison notebook for a live demo — train a medical QA model and compare it against the base model side by side.