Thinking Machines Lab launched its first AI model today, called Inkling. The model is open-weight, meaning any company or developer can download it, run it on their own servers, and modify it however they want. That is a direct contrast to the models from OpenAI, Anthropic, and Google, which are only accessible through a paid service where you send your data to their systems and pay per use.
The company was founded in February 2025 by Mira Murati, who previously served as chief technology officer at OpenAI. It raised $2 billion at a $12 billion valuation in a seed round led by Andreessen Horowitz, with Nvidia, AMD, Cisco, and Jane Street among the investors. A larger fundraise at a reported valuation as high as $50 to $60 billion was discussed last November but did not close. The company's current valuation stands at $12 billion.
Inkling itself is not trying to be the most capable model on the market. The company states that plainly. What it aims to be is a well-rounded, efficient starting point. It processes text, images, and audio together in a single model. It also lets users dial how much thinking time it applies to a problem, so you can trade speed for accuracy depending on what you need.
The real product is not Inkling alone. It is Inkling combined with Tinker, the company's customization platform, where organizations train the model further on their own data. The argument Thinking Machines is making is that a model trained on your specific expertise will outperform a general-purpose model, no matter how powerful. The Bridgewater Associates project is the clearest evidence they have for that claim.
Bridgewater, the world's largest hedge fund, worked with Thinking Machines to train an open model on Bridgewater's own expert-labeled financial decisions. That custom model hit 84.7 percent accuracy on six financial document-sorting tasks. The best general-purpose model tested, with expert prompting applied, reached 78.2 percent. The custom model also cost roughly 14 times less per task to run. Those figures come from the two companies' own testing and have not been independently verified, which is worth keeping in mind.
But the principle is solid, and it is not limited to finance. The same logic applies in insurance claims triage, logistics exception handling, procurement document classification, or any other process where your organization has built up years of specific judgment that no publicly available AI has ever seen.
Microsoft CEO Satya Nadella made a related point this week, in a blog post that drew wide attention given that Microsoft has invested billions in both OpenAI and Anthropic. Nadella argued that companies using proprietary AI models pay twice: once in subscription costs, and again by handing over their internal workflows, corrections, and institutional knowledge, which the model provider can learn from. Every time an employee corrects a wrong AI answer, that correction trains the provider's next model, not yours.
That is the practical argument for owning your own model rather than renting one. It is also the reason large companies such as T-Mobile, SAP, and ADP are reportedly moving toward running open models on their own servers instead of sending queries to external providers.
Thinking Machines is not the only player in this space. Meta's open Llama models have long served this purpose, though Meta's last release was considered underwhelming, and many enterprise teams have been looking for a western alternative to the open models coming from Chinese labs. Inkling is positioned to fill that gap.
The business model here is worth understanding. Inkling is free to download. Thinking Machines does not make money from people who simply use the model. Revenue has to come from Tinker, where organizations pay to customize and host their models, and from the broader service layer around that. That is a different bet than the per-query pricing that OpenAI and Anthropic rely on.
For any organization currently paying monthly for AI tools, the question worth asking now is whether your usage patterns are also feeding your business knowledge into a system your vendor controls. If the answer is yes, that knowledge is worth something, and it is leaving your organization every time you use the tool.