Thinking Machines

Thinking Machines matters because it is making a technically coherent wager that frontier AI will need to be adapted by its users, not only prompted through a hosted interface. Tinker, open model weights, and gigawatt-scale compute put real infrastructure behind that wager. Independent evaluations show a capable first model rather than broad benchmark leadership. The open question is whether the company can sustain the team and execution required to turn that start into a durable alternative to larger labs.^2,5,6,4,16

Official site ↗Models in the Atlas →

Profile

Overview

A frontier team formed around customization

Thinking Machines Lab is a San Francisco AI company founded in 2025 by former OpenAI chief technology officer Mira Murati and a group of researchers and engineers from OpenAI, Meta, and other model labs. It raised a record $2 billion seed round at a reported $12 billion valuation before releasing a product. From the beginning, the company framed its goal around models that people and organizations could understand and customize rather than a single closed assistant delivered on fixed terms.^1,2

Tinker becomes the first product

Its first product, Tinker, launched in October 2025 as a training API for supervised fine-tuning and reinforcement learning on open models. Users control the training loop through a small set of operations while Thinking Machines manages the distributed compute. The company paired the service with research on reproducible inference, efficient adaptation, optimization, and distillation, making customization both a product and the organizing subject of its public technical work.^8,2,9,10

A founding team under pressure

The founding team changed substantially during the lab's first year. Co-founder Andrew Tulloch left for Meta in 2025. In January 2026, Mira Murati said the company had parted ways with co-founder and chief technology officer Barret Zoph; Zoph then joined OpenAI and moved to Google by August. Co-founder Luke Metz also returned to OpenAI. In July, co-founder Lilian Weng left citing health and workload and then rejoined OpenAI. Four of the six original co-founders had therefore left, making retention a material part of any assessment of a company whose initial valuation rested heavily on its concentrated team.^3,14,13

Inkling turns the thesis into a model family

Thinking Machines released Inkling in July 2026, followed by Inkling-Small. Inkling is an open-weight multimodal mixture-of-experts model trained from scratch and designed to be further adapted through Tinker. ARC Prize verified that Inkling was the highest-scoring open-weight model it had evaluated on ARC-AGI-1 and ARC-AGI-2 at the time. Artificial Analysis found a more mixed result on agentic knowledge work, where Inkling trailed Nemotron 3 Ultra and GLM-5.2. A multiyear Nvidia partnership for at least one gigawatt of Vera Rubin systems signals an intention to operate at frontier compute scale, but the lab's clearest current distinction is customizable open models rather than raw benchmark supremacy.^11,7,16,5,6

Reporting and context

We ran Thinking Machines Inkling through our private coding-agent bench. Same harness, same bugs.A direct local evaluation of Inkling that helps separate the model's coding-agent behavior from launch claims and broad aggregate benchmarks.

Notable contributions

01A programmable training API for frontier open modelsTinker exposes forward and backward passes, optimizer steps, sampling, and checkpoint state while the service manages the underlying cluster. The contribution is an unusually direct training interface for researchers, not the invention of fine-tuning or reinforcement learning.^8,2
02Reproducible inference for training workflowsThe lab published methods for reducing nondeterminism in distributed language-model inference, addressing a practical obstacle when training and evaluation depend on repeatable sampled outputs.¹⁰
03Open weights built to be customizedInkling links an open multimodal model family directly to Tinker, so users can inspect the weights and continue training the model on their own tasks. Independent ARC Prize results support capability on one reasoning benchmark, while broader claims still need outside evaluation.^11,7,5
04Real-time interaction model researchThinking Machines published a research preview of models designed to listen, speak, and accept interruption during a shared stream rather than alternating through fixed prompt and response turns.¹²

Sources · 16+

Model lab #19

As of Aug 29, 2026

Inkling validates the customization strategy, but team continuity remains the lab's clearest risk.

Thinking Machines now has more than a financing story. Tinker is a concrete training platform, Inkling leads the open-weight models verified by ARC Prize on two ARC-AGI tests, and the Nvidia agreement supplies a path to frontier-scale compute. Artificial Analysis places Inkling behind leading open models on agentic knowledge work, which argues for a capable-model reading rather than benchmark supremacy. Four of six original co-founders have left, including Lilian Weng in July. The model and product thesis is coherent; the organization still has to show it can retain leadership and execute repeatedly.

Why this editorial standing+

Thinking Machines receives 25 points for model evidence because Inkling leads the open-weight systems verified by ARC Prize on two ARC-AGI tests but trails leading open models on an independent agentic knowledge-work evaluation. Its 17 research-influence points recognize concrete work on reproducible inference, adaptation, and interaction without treating a short publication history as field-wide influence. Tinker and two Inkling releases support 17 points for momentum and 9 for ecosystem development. The Nvidia compute agreement helps trajectory, but four of six original co-founders leaving holds that category to 6.^2,7,16,4

Model evidence: 25 / 35
Research influence: 17 / 25
Release momentum: 17 / 20
Ecosystem: 9 / 10
Trajectory: 6 / 10

Notable people

Mira MuratiCo-founder and chief executivecurrent
BZ
Barret ZophCo-founder and former chief technology officerformer
LW
Lilian WengCo-founderformer
JS
John SchulmanChief scientistcurrent
SC
Soumith ChintalaChief technology officercurrent

Revision history1 updates +

Initial researched dossier
Aug 29, 2026
Established the lab's founding, Tinker strategy, public research, Inkling release, mixed independent model evidence, compute agreement, current leadership, and founding-team changes.