Thinking Machines Lab launches Inkling, an open-weight AI model built for enterprise customization

Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI chief technology officer Mira Murati, released its first proprietary AI model Wednesday morning, called Inkling — an open-weight system that marks a significant departure from the one-size-fits-all approach of larger competitors. The model, which uses a mixture-of-experts architecture with 975 billion total parameters but activates only about 41 billion per task, is designed to be downloaded and modified directly by outside developers and enterprises, positioning it as a flexible alternative to the closed models sold by OpenAI, Anthropic, and Google.

What makes Inkling different from other AI models

Inkling is trained on 45 trillion tokens spanning text, image, audio, and video, and reasons natively across all three modalities, according to the company’s release materials. Unlike flagship models from larger labs that are marketed primarily as general-purpose chatbots, Inkling is designed for organizations that want to adapt AI to their own specific needs. The model includes features such as calibrated responses — flagging uncertainty rather than guessing — and a user-adjustable ‘thinking effort’ dial that trades depth for speed. On one internal benchmark, the company claims Inkling uses a third as many tokens as Nvidia’s Nemotron 3 Ultra to achieve the same coding performance, though the company explicitly states that Inkling is ‘not the strongest model available today, closed or open.’

The strategic bet behind open-weight AI

Thinking Machines Lab is positioning Inkling not as a finished product but as a starting point for enterprise customization. The company’s Tinker platform allows organizations to fine-tune the model for their own data and workflows. This approach is underpinned by a broader argument that centralized AI labs selling the same product to everyone underperform models that organizations can shape themselves. A blog post published by Thinking Machines last week argued that expertise specific to individual organizations is lost when AI is trained centrally and set in stone. The argument is gaining traction: Microsoft CEO Satya Nadella warned in a Sunday blog post that enterprises using proprietary AI models effectively pay twice — once in subscription costs and again by handing over business knowledge embedded in their prompts and corrections, which can be absorbed into future model versions. Hugging Face CEO Clem Delangue made a similar prediction last week, saying frontier models will increasingly be reserved for experimentation while most production AI work shifts to private or open-source alternatives.

Evidence from the Bridgewater Associates project

Perhaps the clearest evidence for this argument comes from a recent project involving Bridgewater Associates, the world’s largest hedge fund. Researchers from both companies took an existing open-source model and trained it further on Bridgewater’s own financial expertise. The result scored 84.7% on financial reasoning tests, beating top proprietary AI models, while costing roughly one-fourteenth as much to run. Those results, published jointly in late June, come from the two companies’ own evaluation, not an independent one, but they illustrate the potential of customized open-weight models.

How Thinking Machines built Inkling — and what it cost

Thinking Machines Lab has emphasized how quickly it brought Inkling to market: roughly nine months from founding to model release, compared with roughly five years for OpenAI and three for Anthropic. The model was trained entirely on Nvidia’s GB300 NVL72 systems, as part of a strategic partnership announced in March that includes deploying a gigawatt of Vera Rubin computing capacity. The company has not disclosed the total cost of training Inkling, nor has it detailed its revenue picture, which by most accounts has not been a primary focus so far. A reported $50 billion fundraising round was said to be coming together last November, though multiple outlets reported it had stalled by January; the company has declined to comment on its funding picture since, though Nvidia said it made a ‘significant investment’ in Thinking Machines when the partnership was announced. The company’s bet may be less that it will eventually spend like its larger rivals than that it won’t need to — because once weights are public, nothing obligates anyone who downloads them to pay Thinking Machines to run them. Revenue, the company says, will come from Tinker, its model-customization platform, via training, fine-tuning, and a cut of the hosting ecosystem built around it.

Distillation and data sourcing questions

Asked whether Inkling was trained on outputs from competitors’ models — a practice known as distillation that has drawn scrutiny industry-wide — the company acknowledged that it partly did. Thinking Machines pretrained Inkling from scratch, but used other open-weight models, including Moonshot AI’s Kimi K2.5, to help generate some of its early post-training data before large-scale reinforcement learning took over. The next model, the company insists, will use fully self-contained post-training instead.

Company culture and headcount

Thinking Machines Lab now employs roughly 200 people, up from levels reported after a wave of departures earlier this year, according to a source close to the company who described the turnover as consistent with a broader industry pattern. The source added that the company is not interested in playing the same talent game as larger rivals, favoring continuity over reliance on any one personality. That stance is notable given how much of the company’s story still runs through its now-famous co-founder, Mira Murati, whether planned or not.

Conclusion

Inkling represents Thinking Machines Lab’s first public proof point after a year and a half of largely behind-the-scenes development. The model is not positioned as best-in-class, but as a well-rounded, customizable alternative to the closed models sold by larger competitors. Whether the company’s bet on open-weight, enterprise-driven AI will succeed depends on whether organizations are willing to invest in fine-tuning their own models rather than relying on centralized labs — and whether Thinking Machines can generate enough revenue from its Tinker platform to sustain its infrastructure ambitions. The broader industry debate over centralized versus customized AI is far from settled, but Inkling gives the argument a tangible, testable form.

FAQs

Q1: What is Inkling?
Inkling is an open-weight AI model released by Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati. It uses a mixture-of-experts architecture with 975 billion total parameters, activating about 41 billion per task, and is designed for enterprise customization through the company’s Tinker platform.

Q2: How does Inkling compare to models from OpenAI, Anthropic, and Google?
Thinking Machines Lab explicitly states that Inkling is not the strongest model available today. It is designed for well-rounded performance and enterprise adaptability rather than top-tier benchmark scores. Unlike flagship models from larger labs, Inkling is open-weight, meaning developers can download and modify it directly.

Q3: How can enterprises use Inkling?
Enterprises can download Inkling and fine-tune it through Thinking Machines Lab’s Tinker platform, which allows customization for specific data and workflows. The company’s revenue model is based on Tinker services — training, fine-tuning, and hosting — rather than on metered access to the model itself.