Features
- Manages distributed training, including GPU scheduling and automatic recovery from checkpoints
- Fine-tunes models with LoRA at a fraction of the compute
- Switches between 1B and 235B MoE models by changing one string of code
- Supports vision-language and large reasoning models like Qwen3-VL and Kimi K2 Thinking
- Includes the open-source Tinker Cookbook for SFT, RLHF and math reasoning
- Samples test prompts from a model while it is still training
- Provides Python API primitives such as forward_backward() and optim_step()
Pricing Details
💵 Pricing & Plans (March 2026 Status): Tinker uses a usage-based pricing model that is significantly more affordable than renting raw A100/H100 clusters. Plan Price (Approx.) Key Features Free Tier $0 Free to start; limited training hours for individual experimentation. Developer Usage-Based Charged per million tokens processed; $0.10/GB-month for storage. Enterprise Custom Dedicated GPU clusters; SOC 2 Type II compliance; 24/7 engineering support. 🎁 How to Get Started: Visit thinkingmachines.ai/tinker. Create an API key in the console, then run pip install tinker. Use the provided Tinker Cookbook to load a base model like Llama-3.2 and start your first training loop using the forward_backward() and optim_step() primitives.