Mira Murati's Thinking Machines Lab debuts Inkling, a 975B-parameter AI model

Thinking Machines Lab released its first AI model Wednesday, betting that organizations will pay to customize open-weight systems rather than rent one-size-fits-all models from OpenAI and Anthropic.

The San Francisco-based startup, founded last year by former OpenAI chief technology officer Mira Murati, unveiled Inkling, a mixture-of-experts model with 975 billion total parameters. The system activates roughly 41 billion parameters per task, a design that keeps inference costs lower than dense models of comparable size. Inkling was trained from scratch on 45 trillion tokens spanning text, images, audio and video, though its current output is limited to text, code and structured data.

"It is not the most performant model available today, closed or open," the company said in a blog post, explicitly setting expectations below frontier systems from OpenAI, Anthropic and Google. "We trained Inkling for solid capabilities across the board rather than state-of-the-art performance in a single area, to serve as a foundation for the models we will train in the future."

Inkling's open-weight release means developers can download, run and fine-tune the underlying system — a contrast to the proprietary, API-only approach of GPT-5, Claude 4 and Gemini 3. The model is available on Tinker, Thinking Machines' customization platform launched last October, and on other developer platforms. On one internal benchmark, Inkling used a third as many tokens as Nvidia's Nemotron 3 Ultra to achieve the same coding performance, though the company did not disclose test conditions for the comparison.

The release tests a central thesis of Thinking Machines: that AI systems organizations can adapt themselves will outperform centrally trained, static models. The company published a blog post last week arguing that centralized labs sell everyone the same product, while enterprises willing to own and customize their models can extract more value. Microsoft chief executive Satya Nadella made a similar argument Sunday, warning that enterprises using proprietary AI effectively pay twice — once in subscription costs and again by handing over business knowledge embedded in prompts and corrections.

The Bridgewater Proof Point

The strongest evidence for Thinking Machines' approach came from a collaboration with Bridgewater Associates, the world's largest hedge fund. Researchers from both companies took Alibaba's open-source Qwen model and trained it further on Bridgewater's proprietary financial expertise. The resulting system scored 84.7% on financial reasoning tests, outperforming top proprietary models while costing roughly one-fourteenth as much to run, according to the companies' own evaluation.

The result mirrors a broader shift in enterprise AI adoption. Hugging Face chief executive Clem Delangue predicted last week 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. Chinese open-source models from Alibaba, DeepSeek and Moonshot AI have already captured significant enterprise adoption in the West, filling the void left after Meta shifted Llama 4 to a proprietary approach.

The Business Model Question

Thinking Machines' revenue model depends on Tinker, not on Inkling itself. Because the model weights are public, users who download them have no obligation to pay Thinking Machines for inference. The company generates revenue through training, fine-tuning services and a cut of the hosting ecosystem built around its models — a structure that differs fundamentally from the metered API access that OpenAI and Anthropic sell.

The company has not disclosed its current revenue or funding picture since a reported $50 billion fundraising round stalled in January. Nvidia made a "significant investment" in Thinking Machines as part of a strategic partnership announced in March to deploy a gigawatt of Vera Rubin computing capacity. Inkling was trained entirely on Nvidia's GB300 NVL72 systems. The company now employs roughly 200 people, up from levels reported after two co-founders departed for OpenAI in January.

For investors, the question is whether Thinking Machines can build a sustainable business on customization revenue alone. OpenAI took roughly five years to bring technology to market and show meaningful revenue; Anthropic took about three. Thinking Machines says it accomplished the same in roughly nine months. But the company's long-term economics remain unproven: if Inkling's open weights mean users never need to pay for inference, the revenue per user will be structurally lower than what API-based competitors generate.

This article is for informational purposes only and does not constitute investment advice.