Thinking Machines releases Inkling, open-weight model focused on enterprise customization

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling on Wednesday morning as its first proprietary AI model. Unlike the flagship models from OpenAI, Anthropic, or Google, Inkling is open-weight, enabling outside developers and companies to download and modify it directly. The model operates as a mixture-of-experts system with 975 billion total parameters, though it draws on only about 41 billion parameters for any given task—a common design that keeps very large models faster and cheaper to run. Inkling was trained on 45 trillion tokens spanning text, image, audio, and video, and reasons natively across all four modalities, though its current generation capabilities are limited to text output, including code, styled artifacts, and structured data.

Thinking Machines released Inkling after a year and a half spent building AI infrastructure largely out of public view. Some of that work surfaced in May as a research preview of "interaction models"—AI designed to listen and speak (and even interrupt) instead of stop and wait as with typical chatbots. The model is designed to give calibrated answers, flagging uncertainty rather than guessing, and lets users dial "thinking effort" up or down to trade speed for thoroughness. On one benchmark, Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra—its latest generation open-weight models—to hit the same coding performance. However, the company explicitly states that Inkling is "not the strongest model available today, closed or open," instead prioritizing well-rounded performance and customizability.

Thinking Machines is marketing Inkling less as a finished product than as a starting point for organizations to fine-tune through Tinker, the company's model-customization platform. The company's central argument, elaborated in a post published last week, is that AI trained centrally by one company and then set in stone underperforms AI that organizations shape themselves because so much expertise is specific to the people who hold it. This thesis has gained broader support: Microsoft CEO Satya Nadella warned in a blog post that enterprises using proprietary models effectively pay twice—once in subscription costs and again by handing over business knowledge embedded in thousands of prompts and corrections, which can be absorbed into future model versions. Hugging Face CEO Clem Delangue predicted that frontier models will increasingly be reserved for experimentation and high-value tasks, while most production AI work shifts to private or open-source alternatives—exactly the split Thinking Machines is building around.

The clearest evidence for Thinking Machines' argument came from a project with Bridgewater Associates, the world's largest hedge fund (which is not a company investor). 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 a fourteenth as much to run, though these results come from the two companies' own evaluation rather than an independent one. Regarding training methodology, Thinking Machines pre-trained Inkling from scratch but used other open-weight models—including Moonshot AI's Kimi K2.5—to help generate some early post-training data before large-scale reinforcement learning took over. The company says the next model will use fully self-contained post-training instead.

Thinking Machines struck a strategic partnership with Nvidia in March to deploy a gigawatt of Vera Rubin computing capacity, and Inkling was trained entirely on Nvidia's GB300 NVL72 systems. The company has not disclosed how it plans to balance training costs against revenue, which multiple outlets reported had become the focus of a $50 billion（約8兆円） fundraising round said to be assembling in November; that round was reported to have stalled by January. Thinking Machines has declined to discuss its funding picture since, though Nvidia said it made a "significant investment" when the companies announced their March partnership. The company's revenue model depends on Tinker—through training, fine-tuning, and a cut of the hosting ecosystem—rather than on metered access to the model weights, since once weights are public, nothing obligates anyone to pay Thinking Machines to run them. Thinking Machines now employs roughly 200 people, up from lower levels after a wave of departures earlier this year, including two co-founders who left for OpenAI in January. The company emphasizes that its culture, by design, favors continuity over reliance on any one personality.

Thinking Machines' release of Inkling reflects a growing conviction in enterprise AI that customization beats standardization. The company's bet rests on the premise that centralized labs like OpenAI and Anthropic sell a one-size-fits-all product that ignores domain-specific knowledge unique to individual organizations. This argument has gained credibility beyond Thinking Machines: Microsoft CEO Satya Nadella recently warned that enterprises using proprietary models effectively pay twice—once in subscription costs and again by handing over business knowledge embedded in thousands of prompts and corrections, which can be absorbed into future model versions. Hugging Face CEO Clem Delangue made a parallel prediction that frontier models will increasingly be reserved for experimentation, while most production work shifts to private or open-source alternatives.

The Bridgewater Associates project offers concrete evidence for this thesis. Researchers from both companies fine-tuned an open-source model on Bridgewater's financial expertise and achieved 84.7% on financial reasoning tests, beating top proprietary models while costing roughly a fourteenth as much to run. This result—though evaluated by the two companies themselves rather than independently—demonstrates the potential value of domain-specific customization. Thinking Machines' speed to market (about nine months) also suggests the company has found a more efficient path than its larger rivals took, though questions remain about whether its spending will eventually scale to OpenAI or Anthropic levels, or whether its efficiency-driven approach fundamentally changes the economics.