Tinker

What it does

Tinker takes in your training data or RL environments, plus a Python training loop that defines your loss function and evaluations, and produces fine-tuned open-weight model checkpoints. Your loop runs on your own CPU-only machine, while Tinker handles distributed training across GPUs and exposes calls such as forward_backward, optim_step, and sample.

Researchers use Tinker for supervised fine-tuning and reinforcement-learning workflows with their own loss functions and evaluations. The training loop remains under their control, while the service manages distributed GPU execution. Supported model choices can be changed through the API, and trained weights can be downloaded for use elsewhere. Training currently uses low-rank adaptation rather than full-model fine-tuning, so teams should evaluate whether that approach fits the intended experiment.

Tinker suits researchers and developers who can write Python and evaluate trained models. Getting started requires an account, API credentials and a local Python environment; training itself runs on managed infrastructure. Charges depend on model, context length, training and sampling activity, with separate checkpoint storage costs. Its serverless inference beta currently supports Inkling and Inkling Small and is not recommended for intensive production use. Downloaded weights can instead be deployed through a suitable external runtime.

What it can do

• Fine-tune open-weight dense and mixture-of-experts models from about 1B to 1T+ parameters
• Run custom training loops with your own loss functions and evaluations
• Perform forward_backward gradient accumulation and optim_step updates on distributed GPUs
• Generate outputs from trained models with sample
• Save and load weights and optimizer state, and download trained weights
• Train vision-language models for image understanding tasks

Example workflows

Train a domain classifier

Illustrative example. Input: a labeled dataset of documents and a Python loss function. Steps: write a training loop calling forward_backward and optim_step, evaluate with sample, save state. Output: a fine-tuned open-weight checkpoint you can download.

Iterate on an RL search agent

Illustrative example. Input: a search environment and reward function. Steps: sample rollouts from the current model, apply policy updates through forward_backward and optim_step, checkpoint progress. Output: a trained agent checkpoint ready for evaluation.

Before you choose

• Training uses LoRA rather than full-model fine-tuning.
• Token charges depend on model and context; checkpoint storage is billed separately.
• Serverless inference is beta, limited to Inkling models and not recommended for intensive production use.

Listed under

• Coding
• Sandboxes and infrastructure
• API
• Developers
• Fine-tuning
• Forecasting
• Model training
• Open models

Get started

Get started via the sign-up page, then follow the docs quickstart.

1. Create an account at the sign-up page and obtain access to Tinker.
2. Install the Python SDK and follow the docs quickstart to configure a training client.
3. Choose a supported model and write a training loop with your data, loss function, and evaluations.

Training, sampling and prefill use model-specific per-million-token rates, which can depend on context length. Cached prefill is discounted relative to ordinary prefill. Checkpoint storage costs $0.10/GB/month. Serverless inference has separate rates and beta restrictions.

Source: thinkingmachines.ai/tinker/pricing, checked 29 Oct 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 29 Oct 2026. Each source supports part of the overview; availability and prices can change.

• Tinker - Thinking Machines Lab ↗thinkingmachines.ai
• Tinker Pricing - Thinking Machines Lab ↗thinkingmachines.ai
• Tinker: a training API for researchers and developers - Tinker Documentation ↗tinker-docs.thinkingmachines.ai

Company website

thinkingmachines.ai: product, pricing and policy pages

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 29 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.