Tinker

by thinking-machines-lab

Training API for researchers: “control every aspect of model training and fine-tuning while we handle the infrastructure.”

See https://thinkingmachines.ai/tinker/. The vault also has a company-level note named tinker under RESOURCES/COMPANIES (not linked here to avoid a same-name clash).

What it is

A Python-first training API that runs LoRA (low-rank adapter) fine-tuning on managed distributed compute. Instead of a one-call fine-tuning job, you write your own training loop and loss locally and Tinker executes the heavy steps.

Core API functions (product page)

• forward_backward: forward and backward pass, accumulating gradients.
• optim_step: update weights from accumulated gradients.
• sample: generate tokens for interaction, evaluation or RL actions.
• save_state: save progress to resume.

Models

30+ open-weight models, including Thinking Machines’ own Inkling and Inkling-Small, Qwen (4B to 397B parameters), NVIDIA Nemotron (30B to 550B), DeepSeek, Moonshot Kimi, Z.AI GLM and OpenAI gpt-oss. The list changes; check the product page.

Availability and billing

Generally available (sign-up at auth.thinkingmachines.ai); large organisations are pointed to enterprise support. Usage-based billing per token plus checkpoint storage; per-model rates are in the docs. The earlier claim “pricing not publicly disclosed / early access” is out of date.

Open-source resources

Tinker Cookbook: https://github.com/thinking-machines-lab/tinker-cookbook (recipes for specialised agents, forecasting, continual learning, research).

Positioning

More control than black-box managed fine-tuning; less infrastructure work than DIY PyTorch/vLLM stacks. Note the LoRA-only approach: full-parameter training is not described on the page.

Not verified

Earlier draft items (dataset versioning, DVC/MLflow integrations, “safety guardrails”, PyTorch/JAX compatibility) were not found on the product page and are omitted.

Sources

• https://thinkingmachines.ai/tinker/ (accessed 2026-09-30)