Thinking Machines Lab

Mira Murati stepped down as OpenAI’s chief technology officer in September 2024; late that year OpenAI’s Barret Zoph and Luke Metz also left OpenAI to co-found Thinking Machines Lab with Murati. As of July 2025, the company was less than a year old.⁠[1][2][3]

The company says its members helped build widely used AI products such as ChatGPT and Character.ai, open-weights models such as Mistral, and open source projects such as PyTorch, OpenAI Gym, Fairseq and Segment Anything, and it plans to frequently publish technical blog posts, papers and code.⁠[4][4]

On July 15, 2025, a company spokesperson confirmed that Thinking Machines Lab had closed a $2 billion seed round led by Andreessen Horowitz, with participation from Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street, at a $12 billion valuation. Before that, the company had struck a deal with Google Cloud to provide computing power for its models.⁠[3][3][3]

The same day the company publicly confirmed its founding team for the first time: CEO Mira Murati, together with John Schulman, who helped build ChatGPT; former OpenAI vice president of research Barret Zoph; Lilian Weng, who worked on AI safety and robotics; Andrew Tulloch, who worked on pretraining and reasoning; and Luke Metz, who worked on post-training. Murati said on X that the company was developing multimodal AI that collaborates with people through conversation, sight and other means, and that its first product would arrive within months, include a significant open source component and serve researchers and startups developing custom models.⁠[1][1][1][1][1]

On September 10, 2025, the company published the first article on its research blog Connectionism, “Defeating Nondeterminism in LLM Inference” (Horace He and others), followed by research articles including Jeremy Bernstein’s “Modular Manifolds,” “LoRA Without Regret” by John Schulman and others, and “On-Policy Distillation” by Kevin Lu and others.⁠[5][5][5][5]

On October 1, 2025, the company launched its first product, Tinker: an API for fine-tuning language models in which researchers control the algorithms and data, while scheduling, resource allocation and failure recovery for distributed training are managed by the company on its internal clusters. Tinker uses LoRA so that multiple training runs share the same pool of compute to lower costs, provides low-level primitives such as forward_backward and sample, and was accompanied by the open-source Tinker Cookbook library of methods. It launched as a private beta, free to start, and its first users included teams at Princeton, Stanford, Berkeley and Redwood Research.⁠[6][6][6][6][6][6][6][6][6]

At launch Tinker supported two families of open source models, Meta’s Llama and Alibaba’s Qwen, including large mixture-of-experts models such as Qwen-235B-A22B; users could fine-tune through supervised learning or reinforcement learning and download the fine-tuned models to deploy themselves. Murati said Tinker makes a capability otherwise reserved for frontier labs accessible to everyone; Schulman said users get full control over the training loop.⁠[7][6][7][7][7][7]

On Friday, October 10, 2025, co-founder Andrew Tulloch announced to employees that he was leaving to join Meta; a company spokesperson confirmed that he “has decided to pursue a different path for personal reasons.” Tulloch had previously worked at OpenAI and Facebook’s AI research group.⁠[8][8][8][8]

On October 29, 2025, the company launched Tinker grants for academia: teaching grants provide $250 in free credits per student for courses using Tinker, and research grants start at $5,000; early recipients included Diyi Yang’s course at Stanford and a deep reinforcement learning course at Carnegie Mellon UniversityCarnegie Mellon UniversityCarnegie Mellon University (CMU) is a private research university in Pittsburgh, Pennsylvania. It grew out of the Carnegie Technical Schools, founded with funding from Andrew Carnegie in 1900, which were renamed the Carnegie Institute of Technology in 1912 and took the present name after merging with the Mellon Institute in 1967. The university is known for computer science, robotics, engineering, and the arts, and has campuses in Qatar, Silicon Valley, and Rwanda. As of September 2025, it enrolled 16,582 students. Its current president is Farnam Jahanian. On September 30, 2026, the university announced a $3 billion gift from Ken Griffin and plans to build a new campus in Miami.Open the full entry. On November 7 the company also issued a call for community projects, planning to regularly showcase users’ projects.⁠[9][9][9][9][10]

On December 12, 2025, Tinker dropped its waitlist and opened to everyone, adding the trillion-parameter reasoning model Kimi K2 Thinking (then the largest model in its lineup), OpenAI API-compatible sampling, and image input through Qwen3-VL models.⁠[11][11][11][11]

On January 14, 2026, OpenAI announced that Thinking Machines Lab co-founders Barret Zoph and Luke Metz were rejoining OpenAI, along with another staff member, Sam Schoenholz. Fidji Simo, OpenAI’s CEO of applications, wrote in an internal memo that Zoph told Murati on Monday of that week that he was considering leaving and was then fired on Wednesday; a reporter wrote on X that Zoph was fired for “unethical conduct,” and a source close to Thinking Machines alleged that he had shared confidential company information with competitors, but WIRED was unable to verify this with Zoph, and Simo said OpenAI did not share Murati’s concerns about Zoph.⁠[2][2][2][2][2][2]

Murati confirmed Zoph’s departure on X and announced that Soumith Chintala would replace him as the company’s chief technology officer.⁠[2]

On March 10, 2026, Thinking Machines Lab and NVIDIA announced a multi-year strategic partnership to deploy at least one gigawatt of next-generation NVIDIA Vera Rubin systems for frontier model training and customizable AI platforms, with deployment targeted for early the following year; the two will also design training and serving systems for NVIDIA architectures. NVIDIA also made a significant investment in Thinking Machines Lab, the amount of which was not disclosed.⁠[12][12][12][12]

On May 11, 2026, the company released a research preview of “interaction models”: models trained from scratch to handle interaction natively, continuously taking in audio, video and text and thinking, responding and acting in real time rather than relying on external scaffolding. The company argues that interactivity should scale alongside intelligence; the preview’s TML-Interaction-Small is a mixture-of-experts model with 276 billion parameters, 12 billion active, and the company acknowledged remaining limitations with very long sessions and network conditions. On May 19 the company announced interactivity research grants of $100,000 each plus $25,000 in Tinker credits.⁠[13][13][13][13][13][14]

On July 15, 2026, the company released Inkling, a model trained from scratch with full open weights: a mixture-of-experts architecture with 975 billion total parameters and 41 billion active, a context window of up to 1 million tokens, pretrained on 45 trillion tokens of text, images, audio and video, able to reason natively over text, images and audio with controllable thinking effort. The company said Inkling is not the strongest model available, but its combination of multimodal capabilities, efficient reasoning and fine-tuning on Tinker makes it a good open-weights base for customization.⁠[15][15][15][15]

Inkling was available for fine-tuning on Tinker at launch (at a 50% discount for a limited time), with full weights uploaded to Hugging Face, and can be called through APIs on Together AI, Fireworks, ModalModalModal is a cloud infrastructure company founded by Erik Bernhardsson and Akshat Bubna. Operated by New York-based Modal Labs, Inc., its platform lets developers configure cloud CPUs, GPUs and containers in code for model inference, training, batch processing and isolated execution of AI agent workloads.Open the full entry, Databricks and Baseten.⁠[15][15][15]

On July 30, 2026, the company released the open-weights model Inkling-Small: 276 billion total parameters and 12 billion active, trained on NVIDIA GB300 NVL72 systems, likewise supporting audio and image reasoning and a 1-million-token context; the company said it achieves performance comparable to Inkling at about a quarter of the size, with output pricing of $1.20 per million tokens (Inkling: $4.05).⁠[16][16][16][16]

On July 31, 2026, the company published an article on the safety of open weights, arguing that once weights are public anyone can use them, that releasing weights indiscriminately is not safe, and that safe release depends both on the model itself and on whether the defensive ecosystem is ready. The company ran internal evaluations of Inkling and Inkling-Small, external testing by four independent organizations and a fine-tuning elicitation study, and concluded that releasing Inkling was not likely to add material risk beyond existing open-weight models. On August 24 the company announced Tinker credit grants of up to $50,000 for safety research on open-weight models.⁠[17][17][17][17][17][18]

On October 15, 2026, the company published a one-year review of Tinker: it launched with 14 text-only models from Qwen and Llama and has since added more than 20 models from Qwen, Nemotron, Kimi, GLM, DeepSeek, GPT-OSS and its own Inkling, ranging from 4 billion to 1 trillion parameters, most supporting image input; the platform added enterprise features such as single sign-on, role-based access control, audit logs and private network connectivity; the company pledges that customer content on Tinker belongs to the customer and is never used for training; and it began offering full fine-tuning as a beta feature.⁠[19][19][19][19][19][19]

Customer examples listed on the Tinker product page include Bridgewater using Tinker to train Qwen3-235B to filter financial information relevant to its investors, Lightning Rod Labs training a model to predict clinical events from patient notes, and Mantic training a model to forecast geopolitical, economic and technological events.⁠[20][20][20]

The company describes itself as an artificial intelligence research and product company. It argues that knowledge of how frontier AI is trained is concentrated in a few top labs, that scientific understanding lags behind advancing capabilities, and that systems are hard to customize to individual needs, so it aims to make AI systems more widely understood, customizable and generally capable. It emphasizes human-AI collaboration rather than solely fully autonomous systems, while prioritizing model intelligence, infrastructure quality and multimodal capabilities, and advocates research and product co-design and an empirical, iterative approach to AI safety.⁠[4][4][4][4][4][4]

Thinking Machines Lab: Official website: https://thinkingmachines.ai/ (opens in a new window).⁠[4]

X: Public account: https://x.com/thinkymachines (opens in a new window)⁠[4]

LinkedInLinkedInLinkedIn is a social networking platform for professionals. It was founded in 2002 by Reid Hoffman and others, launched in 2003, and listed on the New York Stock Exchange in 2011; after Microsoft acquired it for approximately $26.2 billion in 2016, it became a wholly owned Microsoft subsidiary. Its revenue comes from membership subscriptions, advertising, and recruitment solutions.Open the full entry: Public profile: https://www.linkedin.com/company/thinkingmachinesai/ (opens in a new window)⁠[4]

YouTubeYouTubeYouTube is Google’s video-sharing and creator platform, founded in 2005 by Steve Chen, Chad Hurley, and Jawed Karim and acquired by Google in 2006. Built around video uploading, viewing, and creator earnings, the platform has developed Shorts, YouTube Music, Premium, YouTube TV, and products for children, and is headquartered in San Bruno, California, United States.Open the full entry: Public channel: https://www.youtube.com/@thinkymachines (opens in a new window)⁠[4]