Thinking Machines Lab, the AI ​​startup founded by former OpenAI CTO Mira Murati, launched its first in-house AI model Wednesday morning, called Inkling. And unlike flagship models from OpenAI, Anthropic or Google, it is open weight, meaning third-party developers and companies can download and modify it directly.

Inkling is a mixed expert system with 975 billion total parameters, although it only uses a fraction of that (about 41 billion) for any given task, a common design that makes very large models faster and cheaper to run. It was trained on 45 trillion text, image, audio and video tokens, and reasons natively on all four, according to the company’s own release materials. For now, however, its results are limited to text, including code, styled artifacts, and structured data.

The model is Thinking Machines Labs’ first public test after a year and a half spent building AI infrastructure largely out of public view. Some of that work had already emerged in a preview of May research on “interaction models”: AI designed to listen and speak (and even interrupt) rather than stop and wait as typical chatbots do. It’s also proof of the central bet behind the startup, which is that AI that organizations can adapt on their own will outperform the one-size-fits-all models currently sold by larger labs.

Inkling is designed to give calibrated responses, including pointing out uncertainty rather than guessing, and allows users to increase or decrease “thinking effort” when they want to shift for speed. In a benchmark, the company says, Inkling uses one-third as many tokens as Nvidia’s Nemotron 3 Ultra, its next-generation open model, to achieve the same encoding performance.

Thinking Machines does not claim that Inkling is the best in its class. Their most recent blog post explicitly states that Inkling “is not the strongest general model available today, neither open nor closed.” What you are obviously looking for is complete performance.

This raises the question of who, within the enterprise market it is aimed at, this product is really for. Thinking Machines, for now, markets Inkling less as a finished product and more as a starting point, something organizations can refine through Tinker, the company’s model customization platform. This also means that customers, not Thinking Machines, are responsible for ensuring their customizations are secure, for example. (Fine tuning requires great talent to make money with machines.)

OpenAI, Anthropic, and Google have taken a very different approach with ChatGPT, Claude, and Gemini, respectively, which were built to compete first as general-purpose chatbots, with autonomous and agency functions layered on top.

A post published by Thinking Machines last week was clearly intended to be the backdrop for this release. AI that is centrally trained by a company and then set in stone, the company argued in that post, underperforms AI that organizations shape themselves because much of the experience is specific to the people who possess it.

Other arguments against closed models are gaining strength. In a blog post published Sunday, Microsoft CEO Satya Nadella, whose company has invested billions in both OpenAI and Anthropic, warned that companies that use proprietary AI models effectively pay twice: once in subscription costs and again by delivering business insights embedded in their prompts and fixes, which can be absorbed into future versions of the model.

Hugging Face CEO Clem Delangue made a similar prediction in a conversation with TechCrunch last week. Frontier models, he said, will increasingly be reserved for experimentation and high-value tasks, while most AI production work shifts toward private or open source alternatives — the exact division around which Thinking Machines is being built.

The clearest argument for Thinking Machines’ approach came from a recent project with Bridgewater Associates, the world’s largest hedge fund (which, for what it’s worth, is not a Thinking Machines investor). Researchers at both companies took an existing open source model and further trained it with Bridgewater’s own financial expertise. The result was said to score 84.7% on financial reasoning tests, outperforming the best proprietary AI models, while costing about one-fourteenth as much to run, although those results come from the two companies’ own assessment, not an independent one.

Either way, Thinking Machines emphasizes how quickly it got here. It took about five years for OpenAI to bring its technology to market and show revenue, and Anthropic about three. Thinking Machines says it did the same thing in about nine months.

Some may wonder whether Inkling received training on the results of competing models, a practice known as “distillation” that has drawn industry-wide scrutiny. The short answer, according to the company’s own materials, is in part. Thinking Machines pretrained Inkling from scratch, but says it used other open-weight models, including Moonshot AI’s Kimi K2.5, to help generate some of its first post-training data before large-scale reinforcement learning took over. The next model, the company insists, will instead use fully autonomous post-training.

When it comes to costs, Thinking Machines has been more cautious. It partnered with Nvidia in March to deploy a gigawatt of Vera Rubin computing capacity and trained Inkling entirely on Nvidia’s GB300 NVL72 systems, but it hasn’t said how it plans to cover those costs, and revenue, in most cases, hasn’t been a priority. (A $50 billion fundraising round was said to be in the works last November, but had stalled in January; the company has declined to discuss its funding outlook since then.)

A related question is whether Thinking Machines’ spending will one day reach the scale of OpenAI or Anthropic, or whether its efficiency-driven approach means the economy will look different. Put another way, the company’s bet may be less that it will spend like its larger rivals than not need it at all, because once the pesos are public, nothing forces anyone who downloads them to pay Thinking Machines to run them, unlike the metered access that OpenAI and Anthropic sell. It’s Tinker, not the model itself, where the company’s revenue must come from, through training, tuning, and now a part of the hosting ecosystem built around it.

The squad, at least, seems more established. Thinking Machines now employs about 200 people, compared to levels reported after a wave of departures earlier this year, including two co-founders who left for OpenAI in January.

Thinking Machines, for its part, doesn’t seem interested in making individual moves like much of the industry does. According to an internal company source, its culture, by design, favors continuity over dependence on a single personality. It makes sense: It’s a minor setback when people change teams if they were never on a pedestal to begin with. It’s also an extraordinary thing for a company to insist on, given that so much of its own history is still associated with the name of its now famous co-founder, whether she planned it or not.