Tinker

Overview

Tinker is training infrastructure from Thinking Machines Lab for teams that want to fine-tune or run reinforcement learning on large open models without managing GPU clusters. Your code controls the training loop; Tinker runs it on shared GPU pools and handles infrastructure failures.

Main features

• Four core calls: sample (generate tokens for chat, evaluation or RL rollouts), forward_backward (accumulate gradients), optim_step (update weights) and save_state (checkpoint and resume).
• LoRA training: trains a small adapter instead of all weights; Thinking Machines says that with the right setup it matches full fine-tuning while using less compute.
• Recipes: supervised fine-tuning, RL including math and terminal agents, RLHF, on-policy distillation, search tools and forecasting.
• Models: Inkling 2 Small (594B total, 17B active), Inkling (975B, 41B active) and Inkling-Small (276B, 12B active), all with 256K context, plus models such as Kimi-K2.6, GPT-OSS-120B and Qwen3-8B.
• Serverless inference (beta) for the Inkling models.

Pricing

Training, sampling and prefill are priced per million tokens and vary by model and context length; see the Tinker pricing page for current rates. Checkpoint storage costs $0.10 per GB per month.

Who it is for

ML engineers and research teams who want to customise open models with their own data and reward functions, and who prefer writing the training loop themselves.

How to start

Sign up on the Tinker site, pick a small model, and run a supervised fine-tuning recipe on a few hundred examples before moving to RL. Universities and organisations needing wider access can contact the team.

Things to note

You pay for every training and sampling token, so start small and evaluate often. Some older models are being retired, so check the model list before you build on one.

Similar tools

Together AI fine-tuning, OpenAI fine-tuning, Fireworks AI, Hugging Face AutoTrain, Modal.