When I last wrote about Thinking Machines Lab, I compared it to Pied Piper from HBO’s Silicon Valley: a startup backed by extraordinary funding and talent but no public product at the time. The analogy made sense then in July 2025. The startup, founded by former OpenAI CTO Mira Murati, had just raised a record-setting $2 billion seed round at a $12 billion valuation before unveiling what it was actually building.
Since then, Thinking Machines Lab has gone through some changes. Recent reporting by The New York Times describes much of 2025 as a period of internal turmoil for the company, with senior departures, disagreements over technical leadership, and failed acquisition talks, culminating in the exit of several founding researchers. Those exits carried particular weight because the company was originally built around a small circle of prominent OpenAI alumni recruited by Murati, including John Schulman, Barret Zoph, and Luke Metz, as well as former OpenAI research and safety leader Lilian Weng. Zoph and Metz have since returned to OpenAI, while another co-founder, Andrew Tulloch, has moved to Meta. Much of the core team that gave Thinking Machines its early credibility has left.
Mira Murati
The leadership disputes and OpenAI rivalry may make for an engaging Silicon Valley narrative, but they shed little light on the technical direction of the company or the rationale behind Murati’s decision to remain independent. According to the NYT report, senior colleagues urged Murati to pursue a sale, including talks with Meta, but no transaction ever materialized. Murati chose to keep Thinking Machines intact through its first product release, even as the company has navigated upheaval and sought to raise new funding at an aggressive valuation. It’s difficult to guess whether that decision reflects Murati’s confidence in the technology, a need for tighter control during uncertain times, or both, but perhaps it would help to look at what the company has released so far.
Building Tools Instead of Models
Back at the time of the $2 billion seed round, Murati tweeted, “Thinking Machines Lab exists to empower humanity through advancing collaborative general intelligence,” and described how the company was building multimodal AI with an approach centered on open research and software tools designed to help developers and researchers build and customize models. “Soon, we’ll also share our best science to help the research community better understand frontier AI systems,” she wrote.
That first release arrived in October 2025 with Tinker, a platform designed to automate and simplify the fine-tuning of large, open models such as Meta’s Llama and Alibaba’s Qwen families. Rather than debuting a proprietary foundation model, Thinking Machines Lab introduced Tinker as a developer-facing tool aimed at helping researchers, startups, and engineers adapt existing high-end models for specialized tasks without having to manage large GPU clusters or complex training pipelines.
(Kitinut Jinapuck/Shutterstock)
What sets Tinker apart is the way it divides operational responsibility between the user and the system with AI model training and tuning. Developers define their own training loops, loss functions, and evaluation logic in standard Python running locally on CPU machines, while Tinker handles the distributed GPU training required to run those exact computations at scale. This design preserves granular control over supervised fine-tuning and reinforcement learning workflows without requiring users to manage infrastructure, synchronization, or failure recovery. Tinker also allows users to download their trained weights and deploy them wherever they choose, rather than keeping models tied to a single inference service.
Seen in that light, Tinker may help contextualize Murati’s decision to keep Thinking Machines independent. The company’s first product was not an attempt to compete directly with the closed, proprietary systems of its rivals, but an effort to ship infrastructure built around open models, portability, and shared research. At a moment when acquisition talks were underway and some senior researchers were pushing for a sale, Thinking Machines chose to release a tool that stood on its own rather than fold its technology into a larger platform.
Rising Expectations, Stalled Momentum
The release of Tinker did not temper expectations around Thinking Machines Lab so much as accelerate them. In November 2025, roughly a month after the platform’s debut, the company entered discussions about raising a new round at a valuation as high as $50 billion, according to Bloomberg and Reuters. The proposed jump seemed to imply that investors were willing to more than quadruple the startup’s valuation on the strength of its early technical progress and the pedigree of its team.
That round never materialized. By the end of the year, it seems prospective backers could not reconcile the company’s lofty valuation goals with what it had publicly delivered. Tinker may be a technically sophisticated tool, but it’s also an infrastructure product aimed at a relatively narrow audience of AI developers and researchers, instead of a flagship system capable of generating immediate revenue. Without a proprietary foundation model, a defined commercialization strategy, or a clear timeline for larger releases, the gap between expectation and execution has continued to grow.
That tension surfaced openly this month, when internal drama at the company spilled into public view. Murati announced the departure of co-founder and CTO Barret Zoph and named PyTorch co-creator Soumith Chintala as his replacement, calling the change a reset. Within hours, however, Zoph and two other senior researchers had resurfaced at OpenAI, followed by additional resignations in the days that followed. The exits may reflect more than just personnel disputes, instead showing long-running disagreements over product and technical direction and how the company’s $2 billion coffer should be spent. This sudden loss of key staff has turned a stalled funding effort into a larger credibility problem, raising fresh questions about whether the company can still accomplish its goals.
An Uncertain Path Forward
(Sasha Ka/Shutterstock)
As 2026 moves along, Thinking Machines Lab occupies an uncertain position in an AI landscape still defined by hype and extreme expectations, where investor preference seems focused on more established AI players. Tinker is now generally available, with the company lifting its waitlist late last year and expanding the platform to support larger reasoning models, vision-language systems that combine images and text, and OpenAI API–compatible inference. Those updates make Tinker easier to adopt within existing developer workflows, but it remains a specialized tool aimed at AI researchers rather than the mass market.
With no new products announced or even teased, an exodus of talent, and its next funding round unresolved, Thinking Machines Lab now faces the test of aligning its technical vision with a sustainable business plan. Whether the company can stabilize and rebuild its momentum will depend on what it delivers next. Thinking Machines is encountering the same recalibration now happening across the AI sector, as stakeholders begin to reassess what real progress actually looks like beyond shocking headline valuations. After years of enormous capital flowing into frontier AI ahead of products and revenue, investors are increasingly asking harder questions about what gets built, by whom, and at what cost.
The post The Thinking Machines Lab Saga and a Changing AI Market appeared first on AIwire.
About the Author
