Thinking Machines Lab has released Inkling, the first AI model from the startup founded by former OpenAI executives and researchers. The model is open-weight, which means outside researchers and startups can download the model weights and modify them, rather than only renting access through a hosted API.
That choice matters for builders who want more control over how a model behaves, what data it is tuned on, and how much they pay to run it. It also puts Thinking Machines into the busiest lane in AI: big models that can be adapted by outsiders, assuming those outsiders can afford the hardware bill.
Thinking Machines said in a blog post that Inkling was trained from scratch to process text, audio, and video. The company said the model is not at the top of common benchmark tables, which is a refreshingly unglamorous admission in a market where every chart seems to arrive with confetti. Still, Thinking Machines claims Inkling performs strongly across a range of tasks and can handle advanced reasoning and coding work.
There is a catch, and it is not a small one. Inkling has 975 billion parameters, according to Thinking Machines. Models at that scale do not sit comfortably on a developer laptop. The company said Inkling requires a cluster of specialized chips, putting it closer to infrastructure project than weekend download for most users.
A model used to improve itself
Thinking Machines also said it used Inkling during its own fine-tuning process. That reflects a broader pattern in AI development, where companies use existing models to train, evaluate, or improve newer ones. The lab said Inkling helped fine-tune and improve Inkling itself.
The training process produced a behavior the company described in its blog post: the model’s intermediate reasoning explanations became shorter over time. According to Thinking Machines, Inkling reduced some of the grammatical padding in those explanations while keeping them understandable and without changing the final response.
That is the kind of detail model labs tend to present as a window into reasoning. It is also a reminder that “chain of thought” is an artifact produced by a model, not a live feed from a tiny symbolic philosopher in the machine. Thinking Machines’ claim is that this artifact became more compact during training.
Open weights as a strategy
The release fits the public line Thinking Machines has drawn around decentralizing AI development. In a recent blog post, the company argued that AI should not be controlled by a small number of companies and that more people should be able to build models using their own data.
Open-weight models have become attractive because they can be cheaper to operate than closed systems that are typically sold through paid access, and because users can adapt them for specific jobs. According to WIRED, many of the strongest open-weight models now come from China. Thinking Machines says Inkling reaches a similar level of performance.
WIRED reported that Thinking Machines was founded in February 2025 by former OpenAI figures including Mira Murati, who had been OpenAI’s chief technology officer and briefly its chief executive; John Schulman, an OpenAI cofounder involved in the development of ChatGPT; and Lilian Weng, a former OpenAI vice president who worked on safety and robotics.
The company has already attracted unusually large backing. WIRED reported that Thinking Machines received the largest seed funding round on record, valuing the startup at $12 billion at launch. Before Inkling, the company released Tinker, a tool for fine-tuning models, showed a natural voice interaction tool, and published machine-learning research.
Thinking Machines is one of several companies built by former OpenAI staff trying to turn pedigree into a durable AI business. Anthropic, another such company, has become a major competitor with Claude, which WIRED reported has been popular with businesses, especially for coding. WIRED also reported that Anthropic recently filed for an IPO that could value it above $1 trillion.
This story draws on original reporting from WIRED.
