HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

TinkerThinking Machines Lab—TinkerThinking Machines Lab—Thinking Machines Lab—TinkerThinking Machines Lab—

Tinker

by Thinking Machines LabUS

DeployStage 4 of 5

Open to researchers and developers for retraining open-weight models such as Qwen and Llama.

Updated 30 Jul 2026Checked 25 Sep0 updates this week

Milestones

No announced next step
  1. Research blog: making AI answers repeatable10 Sep 2025Complete.
  2. Tinker private beta1 Oct 2025Complete.

Most important updates

  • 30 Jul 2026
  • 15 Jul 2026
  • 12 Dec 2025
  • 1 Oct 2025

Current obstacles

  • Standing outClouds and AI labs also offer fine-tuning. Tinker must win on flexibility for researchers.

Physics limits

  • Fine-tuning adds skills more than knowledgeLow-rank adapters change a small sliver of the weights (often under 1%). They steer format and skills well but store little new factual knowledge compared with pretraining.
  • Learning new things erodes old onesTraining on a narrow task can overwrite general abilities ('catastrophic forgetting'), because the same numbers store many skills at once.
  • Only checkable skills train wellReasoning is learned by rewarding answers a program can grade, such as maths and code. Where no automatic checker exists (strategy, taste, open research) the signal is weak.

How it works

4 parts
Cabled servers in a rack: Tinker runs your fine-tuning loop on its own GPU clusters, so you never manage machines like these
Cabled servers in a rack: Tinker runs your fine-tuning loop on its own GPU clusters, so you never manage machines like thesePhoto: Tyler · Unsplash License (opens unsplash.com)
API

You write the loop

Researchers write training code as simple Python calls (compute gradients, update, sample) while Tinker runs it on its own GPU clusters.

LoRA

Small add-on weights

Tinker trains LoRA adapters, small low-rank matrices added to a frozen base model, so many users can share the same GPUs cheaply.

Methods

Examples or rewards

It supports learning from examples and reinforcement learning from rewards, on open models like Qwen, Llama and Thinking Machines' Inkling.

Ownership

Take your weights

Users can download the trained adapter weights and run them anywhere, rather than being locked into Tinker's servers.

Update log

4 updates

Thu 30 Jul

  • Minor: BlogSoftware

Wed 15 Jul

  • Major: BlogSoftware

Fri 12 Dec 2025

  • Minor: BlogSoftware

Wed 1 Oct 2025

  • Minor: BlogSoftware

About Thinking Machines Lab

The team behind Tinker

Thinking Machines Lab

Tinker fine-tuning service, AI research

Thinking Machines Lab was founded in February 2025 by Mira Murati, OpenAI's former CTO, with researchers from OpenAI, Meta and Mistral. It publishes research and offers Tinker, a service for customising open-weight models.

  • Raised a $2B seed round at a $12B valuation led by Andreessen Horowitz in July 2025, one of the largest ever.
  • Launched Tinker in October 2025 so researchers can customise open models without managing GPUs.
  • Co-founder and CTO Barret Zoph left for OpenAI in January 2026.

thinkingmachines.ai (opens thinkingmachines.ai)Company profile →

Founded
20251 yrs
Headquarters
United States
Status
Private
Valuation
$12BprivateJul 2025
Raised
$2B1 round
Last round
Seed · $2BJul 2025$12B post
Works in
AIFrontier models
Coverage
1 program · 6 updateslatest 30 Jul 2026checked 25 Sep
Lead investors
Andreessen Horowitz
People
Mira MuratiFounder and CEOJohn SchulmanCo-founder and Chief Scientist
Primer