Tinker Documentation
by Thinking Machines Lab
Write your own LLM training loop locally while Tinker handles the GPU cluster.
Overview
Tinker is described in its own docs as "a cloud training API that gives you full control over your data, algorithms, and models. Write training loops locally — Tinker handles the GPUs." It supports model-agnostic LoRA training from 1B to 1T+ parameters across dense and mixture-of-experts architectures, including text and vision models, and covers SFT, RL (GRPO, PPO), DPO, and distillation. The documentation is organized into the Tinker SDK (quickstart, models & pricing, data model, LoRA primer), a Loss Functions section (cross-entropy, importance sampling, PPO, CISPO, DRO, and custom losses), a Cookbook (supervised learning, reinforcement learning, evaluation, storage), numbered Tutorials running 101 through 504, and a full API reference. It advertises 28+ supported models — spanning Thinking Machines' own Inkling, NVIDIA Nemotron, Moonshot Kimi, Alibaba Qwen, OpenAI GPT-OSS, and DeepSeek families — plus 15+ training recipes covering math RL, code RL, tool use, and multi-agent training. The docs and tutorials are free to read; running training is billed per million tokens across prefill, sample, and train operations (with an 80% discount on cached prefill tokens and checkpoint storage at $0.10 per GB per month), and MoE models are priced by active parameters.
At a Glance
- Topic
- Fine-Tuning
- Level
- Advanced
- Format
- Documentation
- Cost
- Paid
- Duration
- Self-paced; 23 interactive tutorials + cookbook
- Provider
- Thinking Machines Lab
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Build custom LLM training loops from four primitives: forward_backward, optim_step, save_weights_and_get_sampling_client, and sample
- ✓Run supervised fine-tuning, DPO, distillation, and RL (GRPO, PPO, CISPO) against open-weight models without managing GPUs
- ✓Apply LoRA correctly — rank, target modules, and when low-rank adaptation matches full fine-tuning — via the LoRA primer
- ✓Pick and implement the right loss function for a post-training objective, including custom losses
- ✓Follow end-to-end recipes for math RL, code RL, tool use, and multi-agent training
- ✓Reason about fine-tuning cost: per-token prefill/sample/train pricing, cached-prefill discounts, and MoE active-parameter economics
Highlights
- •Four-primitive API keeps the training algorithm in your Python file while distributed scheduling, allocation, and GPU-crash recovery happen server-side
- •Switching from a small model to a large MoE is a one-string change
- •Cookbook abstractions sit on top of the raw loop, so you can start high-level and drop down
- •Covers RL post-training in production detail, not just SFT
Who It’s For
Best For
- ✓ML engineers doing post-training or RL on open-weight LLMs
- ✓Researchers who need algorithmic control but not a GPU cluster
- ✓Teams evaluating LoRA fine-tuning versus prompting or RAG
- ✓AI engineers building domain-specialized models on Qwen, Kimi, GPT-OSS, or DeepSeek
Prerequisites
- •Strong Python and PyTorch-style training intuition
- •Familiarity with transformer fine-tuning concepts (LoRA, SFT, RLHF/GRPO)
- •A Tinker API account (usage is billed per token)
FAQ
What is Tinker Documentation?
Official documentation for Tinker, Thinking Machines Lab's cloud training API for fine-tuning open-weight LLMs with LoRA. Written for ML engineers and researchers who want direct control over the loss function, data, and RL loop instead of a black-box fine-tuning endpoint.
Is Tinker Documentation free?
Tinker Documentation is a paid resource.
What level is Tinker Documentation for?
Tinker Documentation is aimed at a advanced audience. Recommended background: Strong Python and PyTorch-style training intuition, Familiarity with transformer fine-tuning concepts (LoRA, SFT, RLHF/GRPO), A Tinker API account (usage is billed per token).
How long does Tinker Documentation take?
Expect roughly Self-paced; 23 interactive tutorials + cookbook. Most learners work through it at their own pace.
What will I learn from Tinker Documentation?
You'll learn: Build custom LLM training loops from four primitives: forward_backward, optim_step, save_weights_and_get_sampling_client, and sample; Run supervised fine-tuning, DPO, distillation, and RL (GRPO, PPO, CISPO) against open-weight models without managing GPUs; Apply LoRA correctly — rank, target modules, and when low-rank adaptation matches full fine-tuning — via the LoRA primer; Pick and implement the right loss function for a post-training objective, including custom losses; Follow end-to-end recipes for math RL, code RL, tool use, and multi-agent training; Reason about fine-tuning cost: per-token prefill/sample/train pricing, cached-prefill discounts, and MoE active-parameter economics.