Tinker

Founded by Mira Murati in 2025

Fine-tune and train open-source language models using simple API functions without managing GPUs

Pricing model

Pay As You Go

Rating

★ People love it

Time to value

Quick Setup (< 1 hour)

You can use Tinker to train and fine-tune open-source large language models without managing cloud infrastructure or GPU clusters. The service provides a programmatic interface with basic training functions to perform forward and backward passes, update weights, sample model outputs, and save checkpoints. It specializes in LoRA fine-tuning and reinforcement learning setups across models like Qwen, Nemotron, and DeepSeek. You maintain complete control over your datasets, loss metrics, and reward algorithms while Tinker handles compute scheduling and distributed hardware execution behind the scenes.

What Tinker does

Accumulate gradients across custom training batchesUpdate model adapter weights with optimizer stepsSample completion tokens to test model responsesSave and export trained model checkpoint weightsConfigure LoRA ranks and adapter parametersRun off-policy distillation experimentsFine-tunes open-source models using parameter-efficient LoRA adaptersProvides programmatic primitives for forward passes, backward passes, and weight updatesSamples tokens directly from training checkpoints for evaluation or reinforcement learningSaves training checkpoints to persistent storage with automatic resumptionOrchestrates distributed training clusters across high-performance GPUs automaticallySupports leading open architectures including Qwen, Nemotron, and DeepSeekEnables custom reinforcement learning loops and direct preference optimization