Thinking Machines Unveils Inkling, Its First Open Model, Doubling Down Against Universal AI

Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI CTO Mira Murati, unveiled its inaugural proprietary AI model, Inkling, on Wednesday morning. A significant departure from the flagship offerings of OpenAI, Anthropic, or Google, Inkling is an open-weight model, providing external developers and organizations the freedom to download and directly modify its core components.

Inkling is architected as a mixture-of-experts system, boasting a formidable 975 billion total parameters. However, for any specific task, it intelligently activates only a fraction of these — approximately 41 billion — a design choice common among large models to ensure faster and more cost-efficient operation. According to the company's official release materials, the model underwent training on an extensive dataset of 45 trillion tokens encompassing text, image, audio, and video, enabling it to reason natively across all these modalities.

This release marks the company's first public demonstration of progress after a year and a half dedicated to developing AI infrastructure largely out of the public eye. Portions of this foundational work previously surfaced in a May research preview of "interaction models" — AI systems engineered to actively listen, speak, and even interrupt, unlike conventional chatbots that require users to wait. Inkling also serves as a critical test for Thinking Machines' core hypothesis: that AI solutions tailored and adapted by individual organizations will ultimately outperform the generic, "one-size-fits-all" models currently offered by leading AI labs.

Inkling stands out for its design philosophy, which prioritizes calibrated responses. It is engineered to explicitly flag uncertainty rather than venture guesses and allows users to adjust its "thinking effort" to balance speed against precision. In a disclosed benchmark, the company claims Inkling achieves the same coding performance as Nvidia’s Nemotron 3 Ultra while consuming only a third of the tokens. Importantly, Thinking Machines openly states that Inkling is not positioned as a best-in-class model; its briefing materials explicitly acknowledge it is “not the strongest model available today, closed or open.” Instead, its objective is to deliver robust, well-rounded performance.

This strategic positioning naturally prompts inquiry into Inkling's target audience, which is clearly the enterprise sector. Thinking Machines is currently marketing Inkling not as a final product, but rather as a foundational framework for organizations to fine-tune via Tinker, its dedicated model-customization platform. This approach contrasts sharply with the strategies of OpenAI, Anthropic, and Google, whose flagship products like ChatGPT, Claude, and Gemini were initially developed as general-purpose chatbots, with more advanced agentic and autonomous features subsequently layered on.

A post published by Thinking Machines the previous week provided the theoretical underpinnings for this launch. In it, the company contended that AI centrally trained by a single entity and then deployed as a static solution will inevitably underperform AI that organizations actively shape themselves, primarily because much valuable expertise is inherent to the specific individuals and contexts within an organization. The overarching argument is that while centralized labs offer a standardized, repeatedly refined product, enterprises willing to customize and own their AI models can extract significantly greater value.

This argument is steadily gaining momentum within the industry. In a blog post published on Sunday, Microsoft CEO Satya Nadella, whose company has invested billions in both OpenAI and Anthropic, cautioned that enterprises utilizing proprietary AI models effectively incur a double cost: once through subscription fees, and again by ceding valuable business knowledge embedded within their extensive prompts and corrections, which can then be assimilated into future model iterations.

Hugging Face CEO Clem Delangue echoed a similar sentiment in a recent conversation with TechCrunch. He predicted that frontier models would increasingly be reserved for experimental applications and highly specialized, high-value tasks, while the bulk of production AI work would transition toward private or open-source alternatives — precisely the market dynamic Thinking Machines aims to capitalize on.

Compelling evidence supporting this perspective recently emerged from a collaboration with Bridgewater Associates, the world's largest hedge fund (and notably, not an investor in Thinking Machines). Researchers from both entities took an existing open-source model and further trained it using Bridgewater's proprietary financial expertise. The resulting model achieved an impressive 84.7% on financial reasoning tests, surpassing leading proprietary AI models, while costing approximately one-fourteenth as much to operate. It is worth noting, however, that these results, jointly published in late June, stem from the two companies’ own evaluation, not an independent assessment.

Thinking Machines has also highlighted its rapid development timeline. While OpenAI took roughly five years and Anthropic approximately three years to bring their technology to market and demonstrate revenue, Thinking Machines asserts it achieved comparable milestones in just about nine months.

The question of whether Inkling was trained using outputs from competitor models, a practice known as distillation that has faced industry scrutiny, is also addressed. According to the company's materials, the short answer is "partly." Thinking Machines pretrained Inkling from scratch, but it did utilize other open-weight models, including Moonshot AI’s Kimi K2.5, to assist in generating some of its early post-training data before large-scale reinforcement learning processes took over. The company assures that its next model will employ a fully self-contained post-training methodology.

Regarding its cost structure, Thinking Machines has maintained a more discreet stance. The company established a strategic partnership with Nvidia in March to deploy a gigawatt of Vera Rubin computing capacity, and Inkling itself was trained entirely on Nvidia’s GB300 NVL72 systems. However, the company has not yet detailed its plans for balancing these substantial investments against revenue, which, by most accounts, has not been a primary focus to date. A reported $50 billion fundraising round, said to be in progress last November, was subsequently reported by multiple outlets to have stalled by January. The company has since declined to comment on its funding status, although Nvidia confirmed a “significant investment” in Thinking Machines when the partnership was announced in March.

A related consideration is whether Thinking Machines' expenditure will ever scale to match that of OpenAI or Anthropic, or if its efficiency-driven approach necessitates a fundamentally different economic model. In essence, the company's strategic wager may not be that it will eventually spend like its larger competitors, but rather that it won't need to at all. Once model weights are public, there is no obligation for those who download them to pay Thinking Machines for their operation, unlike the metered access sold by OpenAI and Anthropic. Consequently, the company's revenue generation must primarily derive from Tinker — through training, fine-tuning services, and, moving forward, a share of the hosting ecosystem that develops around its models.