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://thinkingmachines.ai/tinker/ recorded 2026-08-09

  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.

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

  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://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://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://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.

Thinking Machines' own Tinker page names Glean/Waldo, Chroma, Lightning Rod Labs, Mantic, Trajectory, MIT, Stanford and Axiom Math among users of the API. Level 2 on named adopters rather than a download or seat count, none being published for a hosted training API.

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

  Case-study section now names Glean (Waldo), Chroma (Context-1), Lightning Rod Labs, Mantic, and Trajectory as customers, plus research collaborations credited to MIT, Stanford, and Axiom Math; no aggregate user count, download count, or revenue figure appears anywhere on the page.

Tinker hides distributed multi-GPU training behind a four-function API, covering LoRA training plus sampling for reinforcement learning and evaluation, across a wide model range from 1B up to large mixture-of-experts checkpoints. The largest it lists is NVIDIA Nemotron-3-Ultra-550B-A55B-BF16, at 550B total and 55B active parameters. It stays one band below the top because the service is LoRA-centric, with no full fine-tuning and no full RL stack, and because the judgment rests on that feature comparison rather than on any MLPerf Training result.

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

  FAQ still states "Tinker is a flexible API for efficiently fine-tuning open source models with LoRA." Current model tiles top out at NVIDIA Nemotron-3-Ultra-550B-A55B-BF16; no full-fine-tuning or full-RL-stack option is described anywhere on the page.

• https://tinker-docs.thinkingmachines.ai/tinker/models/ recorded 2026-08-09

  Retired Models page lists a June 12, 2026 retirement batch including Qwen3-235B-A22B-Instruct-2507 and six Llama checkpoints (Llama-3.3-70B-Instruct, Llama-3.1-70B, Llama-3.1-8B, Llama-3.1-8B-Instruct, Llama-3.2-3B, Llama-3.2-1B), confirming the model roster the score's value field cites has shifted since it was last read.