Thinking Machines Lab, an AI startup established by former OpenAI CTO Mira Murati, unveiled its initial in-house AI model on Wednesday morning, named Inkling. Unlike the primary models from OpenAI, Anthropic, or Google, it boasts open weights, allowing external developers and companies to download and modify it directly.
Inkling operates as a mixture-of-experts system with a total of 975 billion parameters, though it utilizes only about 41 billion for specific tasks, which is a common design promoting faster and more cost-effective operation for large models. The model was trained on 45 trillion tokens encompassing text, images, audio, and video, and it reasons natively across all four types, as per the company’s release documents. However, its outputs are currently confined to text, which includes code, styled outputs, and structured data.
This model represents Thinking Machines Labs’ first public demonstration following a year and a half of constructing AI infrastructure mostly out of public sight. Some aspects of this work were highlighted in a May research preview showcasing “interaction models” — AI crafted to engage in dialogue rather than pause and wait, as traditional chatbots do. It also serves as a test of the startup’s core premise that AI, which organizations can tailor to their needs, will surpass the generic models offered by the largest labs.
Inkling is engineered to provide calibrated responses, acknowledging uncertainty instead of making guesses, and allows users to adjust “thinking effort” to expedite the response when necessary. According to the company, in one benchmark test, Inkling utilizes one-third the tokens of Nvidia’s Nemotron 3 Ultra — its latest generation open-weight model — while achieving equivalent coding performance.
Thinking Machines does not assert that Inkling is the top model available. Its latest blog post clearly states that Inkling is “not the strongest overall model available today, open or closed.” Instead, the focus appears to be on achieving balanced performance.
This raises the question of which segment of the enterprise market the product truly targets. Thinking Machines is presently positioning Inkling more as an initial resource rather than a finalized product, intending for organizations to refine it themselves through Tinker, the company’s model-customization platform. As such, customers bear the responsibility for ensuring their customizations are secure, which requires substantial machine-learning expertise.
OpenAI, Anthropic, and Google have all adopted a notably different strategy with ChatGPT, Claude, and Gemini, respectively, which were primarily developed as general-purpose chatbots with agentic, autonomous functionalities added later.
A post issued by Thinking Machines the previous week appeared aimed to provide context for this release. The company contended in that post that AI centrally trained by one firm and then made static falls short compared to AI shaped by organizations, as much of the expertise is unique to those who possess it.
Criticism of closed models is growing stronger. In a blog post released on Sunday, Microsoft CEO Satya Nadella — whose company has poured billions into both OpenAI and Anthropic — cautioned that enterprises utilizing proprietary AI models effectively incur costs twice: first through subscription fees and secondly by surrendering business knowledge embedded in their prompts and corrections, which can be utilized in future model iterations.
Hugging Face CEO Clem Delangue echoed a similar sentiment in a conversation with TechCrunch the previous week. He suggested that frontier models will increasingly be designated for experimentation and high-value endeavors, while the majority of production AI tasks shift toward private or open-source alternatives — the precise focus that Thinking Machines is developing around.
The most compelling endorsement for Thinking Machines’ strategy emerged from a recent collaboration with Bridgewater Associates, the world’s largest hedge fund (which, for what it’s worth, is not an investor in Thinking Machines). Researchers from both firms took an existing open-source model and enhanced it further utilizing Bridgewater’s financial expertise. The outcome reportedly achieved a score of 84.7% on financial reasoning assessments, surpassing leading proprietary AI models, while operating at around one-fourteenth the cost — although these results stem from the evaluation conducted by the two companies, not from an independent source.
Regardless, Thinking Machines is highlighting how rapidly it has progressed. OpenAI required approximately five years to commercialize its technology and produce revenue, while Anthropic took around three. Thinking Machines asserts it accomplished the same feat in about nine months.
Some may question whether Inkling was trained using outputs from competitors’ models, a practice referred to as “distillation,” which has attracted scrutiny throughout the industry. The concise answer, according to the company’s literature, is partly. Thinking Machines initially trained Inkling from scratch, but it claims to have used 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 took precedence. The company insists that its subsequent model will rely on fully self-contained post-training methods.
On the financial front, Thinking Machines has been more reticent. It formed a partnership with Nvidia in March to deploy a gigawatt of Vera Rubin computing power and trained Inkling entirely on Nvidia’s GB300 NVL72 systems — but has not disclosed how it plans to manage those expenses, and, according to most reports, revenue hasn’t been a primary focus. (A rumored $50 billion fundraising round was reportedly in progress last November but had stalled by January; the company has refrained from discussing its funding status since.)
Another pertinent question is whether Thinking Machines’ spending will ever rival the scale of OpenAI’s or Anthropic’s, or if its efficiency-oriented approach indicates a different economic picture. To put it another way, the hypothesis put forth by the company may be that it will not need to spend in line with larger competitors at all — because once weights are publicly available, there are no obligations for anyone who accesses them to compensate Thinking Machines for their use, unlike the subscription-based access offered by OpenAI and Anthropic. It’s Tinker, rather than the model itself, from which the company aims to generate revenue, via training, fine-tuning, and, now, a share of the hosting ecosystem built around it.
Employee count, at least, appears to be more stabilized. Thinking Machines now employs roughly 200 individuals, an increase from the levels reported following a wave of departures earlier this year, including two co-founders who transitioned to OpenAI in January.
Thinking Machines, for its part, does not seem inclined to highlight individual maneuvers as much of the industry does. According to an insider source, the company’s culture, by design, prioritizes continuity over reliance on any singular persona. This approach makes sense: it’s less disruptive when personnel shift teams if they were never elevated to a prominent status initially. It’s also quite notable for a company to adhere to this principle, considering how much of its narrative is still tied to the name of its now-prominent co-founder, whether intentional or otherwise.
