Tinker Documentation
by Thinking Machines Lab
Write the training loop yourself and let someone else's GPUs run it.
Overview
Tinker is a cloud training API from Thinking Machines Lab: you write the training loop on your own machine and the service executes it on managed GPU clusters, so distributed training is abstracted away without the algorithm being hidden from you. The documentation is organised into the Tinker SDK reference, the Cookbook, a Changelog, Support, and tutorials graded Basics, Core Concepts, Cookbook Abstractions, Advanced and Deployment. The SDK exposes three clients (ServiceClient, TrainingClient, SamplingClient) and a handful of primitives that map directly onto one training step: forward_backward computes gradients, optim_step applies an Adam update, save_weights_and_get_sampling_client persists a checkpoint, and sample generates from it. Everything is LoRA-based, and a dedicated LoRA Primer argues that for supervised fine-tuning on small-to-medium instruction and reasoning datasets LoRA performs the same as full fine-tuning. The Models and Pricing page lists Tinker IDs for Qwen3.5/3.6/3.8, NVIDIA Nemotron-3, Kimi-K2.6, gpt-oss-20b and 120b, DeepSeek-V3.1 and the lab's own Inkling models, with per-million-token rates split across three meters (prefill, sample, train), checkpoint storage at $0.10 per GB per month, and an 80% discount on cached prefill tokens. MoE models are priced by active parameters. The companion tinker-cookbook repository (Apache-2.0, roughly 4.1k stars) carries runnable recipes for chat SFT, math and code RL, DPO and RLHF preference learning, distillation, tool use, multi-agent RL and VLM image classification. Tinker left its waitlist and reached general availability on 12 December 2025, adding vision input via Qwen3-VL and an OpenAI-compatible inference interface.
At a Glance
- Topic
- Fine-Tuning
- Level
- Advanced
- Format
- Documentation
- Cost
- Freemium
- Duration
- ~6-10 hours to work through the tutorials and cookbook; reference docs are ongoing
- Provider
- Thinking Machines Lab
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Write a supervised fine-tuning loop using forward_backward and optim_step directly
- ✓Decide when LoRA is sufficient and when full fine-tuning is warranted
- ✓Run reinforcement learning post-training with GRPO, PPO and DPO recipes
- ✓Estimate training cost from the prefill, sample and train token meters
- ✓Checkpoint a run and sample from it through an OpenAI-compatible interface
- ✓Distil a large teacher model into a smaller deployable student model
- ✓Adapt the cookbook's multi-agent RL and tool-use recipes to your dataset
Highlights
- •Exposes the actual training primitives instead of a one-click fine-tune button
- •The same code path covers a 4B dense model and a trillion-parameter MoE like Kimi-K2.6
- •Publishes per-model, per-meter pricing — unusual for a hosted training service
- •The Apache-2.0 tinker-cookbook is effectively the syllabus: every concept has runnable code
- •Vision fine-tuning via Qwen3-VL and OpenAI-compatible sampling shipped at GA in December 2025
Who It’s For
Best For
- ✓Research engineers running RL or DPO post-training without a GPU cluster
- ✓ML teams that need algorithmic control but not infrastructure ownership
- ✓Practitioners benchmarking LoRA against full fine-tuning on real workloads
- ✓Anyone budgeting a fine-tuning project who needs concrete token prices
Prerequisites
- •Solid Python and PyTorch-level familiarity with training loops and optimizers
- •Understanding of tokenization, loss functions and learning-rate scheduling
- •A Tinker API key; the service is usage-priced per million tokens
FAQ
What is Tinker Documentation?
The official documentation for Tinker, Thinking Machines Lab's cloud training API for LoRA fine-tuning of open-weight models. It covers the low-level primitives forward_backward, optim_step, sample and save_weights, plus a LoRA primer, tutorials graded from basics to deployment, a models-and-pricing table, and the open-source Tinker Cookbook of SFT, RL, DPO and distillation recipes. Aimed at engineers who want post-training control without owning distributed GPU infrastructure.
Is Tinker Documentation free?
Tinker Documentation offers free content, with paid options for certificates or premium features.
What level is Tinker Documentation for?
Tinker Documentation is aimed at a advanced audience. Recommended background: Solid Python and PyTorch-level familiarity with training loops and optimizers, Understanding of tokenization, loss functions and learning-rate scheduling, A Tinker API key; the service is usage-priced per million tokens.
How long does Tinker Documentation take?
Expect roughly ~6-10 hours to work through the tutorials and cookbook; reference docs are ongoing. Most learners work through it at their own pace.
What will I learn from Tinker Documentation?
You'll learn: Write a supervised fine-tuning loop using forward_backward and optim_step directly; Decide when LoRA is sufficient and when full fine-tuning is warranted; Run reinforcement learning post-training with GRPO, PPO and DPO recipes; Estimate training cost from the prefill, sample and train token meters; Checkpoint a run and sample from it through an OpenAI-compatible interface; Distil a large teacher model into a smaller deployable student model; Adapt the cookbook's multi-agent RL and tool-use recipes to your dataset.
Topics
Sources
This page was written from 5 sources, 3 on domains other than tinker-docs.thinkingmachines.ai.