Tinker

license

Proprietary

source

closed

engine

closed(hosted distributed-training service on TML infra)

cookbook

open(Apache-2.0, thinking-machines-lab/tinker-cookbook)

api

public-but-proprietary(forward_backward/optim_step/sample/save_state)

method

LoRA(+RL via sample)

access

paid+account-gated

The core training service is a proprietary hosted API, and only the example cookbook is OSI-licensed, so the product is closed. It is not open core either, because no self-hostable engine exists to be the core.

• https://github.com/thinking-machines-lab/tinker-cookbook recorded 2026-08-09

  GitHub repository page for thinking-machines-lab/tinker-cookbook, described in its own metadata as 'Post-training with Tinker' - example and recipe code that calls the hosted Tinker API rather than the training engine itself.

• https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/LICENSE recorded 2026-08-09

  Full text of the Apache License, Version 2.0 - the tinker-cookbook repository's own license, unchanged from the prior read (fetched via the raw.githubusercontent.com redirect).

• https://thinkingmachines.ai/legal/terms/ recorded 2026-09-23

  The Service Terms of Use (last updated January 10, 2026) grants Authorized Users "a non-exclusive, limited, non-transferable, non-sublicensable, worldwide right to access and use the Services" - a proprietary terms-of-use grant over the hosted platform, not an open-source license to the underlying engine.

• https://thinkingmachines.ai/news/ recorded 2026-08-09

  News archive lists "Oct 1, 2025 Announcing Tinker" as its own dated entry, confirming the product's launch date cited in prose.

• https://thinkingmachines.ai/tinker/ recorded 2026-09-23

  Describes Tinker as "a training API for researchers" where users "Control every aspect of model training and fine-tuning while we handle the infrastructure." Offers sign-up/sign-in, docs, and per-token/per-GB-month pricing; no self-host or download path exists for the training engine itself, only a link out to the separate tinker-cookbook example repo.

Thinking Machines names companies and research groups such as Glean, Chroma, MIT and Stanford as users of the API on its own product page. It publishes no download, seat or user count for the hosted service, so named adopters are the only public evidence of use.

• https://thinkingmachines.ai/tinker/ recorded 2026-09-28

  Case-study section names Glean ("Waldo is an agentic search model that hands the reasoning model the tool plan it needs for the job, fast."), Chroma, Lightning Rod Labs, Mantic and Trajectory, plus research quotes from MIT, Stanford and Axiom Math ("AxiomProver is an AI mathematician that conjectures, proves, and verifies its own work in Lean."). Searched the body for user count, download count, seat count and revenue: none.

Tinker hides distributed multi-GPU training behind four functions and supports LoRA training and RL sampling on open-weight models, up to very large mixture-of-experts checkpoints. The service is LoRA-only, with no full fine-tuning or full RL stack, which keeps it behind frameworks such as Megatron-LM and verl that train entire models at scale.

• https://thinkingmachines.ai/tinker/ recorded 2026-09-28

  FAQ states "Tinker is a flexible API for efficiently fine-tuning open source models with LoRA." and argues that "LoRA matches the learning performance of full fine-tuning". The supported-models list runs from "Qwen3.5-4B" up to "Nemotron-3-Ultra-550B-A55B-BF16". Searched the body for a full fine-tuning or full RL stack offering: the only full fine-tuning mention is that LoRA comparison.

• https://tinker-docs.thinkingmachines.ai/tinker/models/ recorded 2026-09-28

  Model table reads "29 models shown" and lists Inkling at 975B total parameters and "Nemotron-3-Ultra-550B-A55B" at 550B, with a ":peft:" suffix on the 256K-context Tinker IDs, plus a Retired Models section. Searched the body for full fine-tuning: none.