On Wednesday morning, Thinking Machines Lab, an AI firm created by former OpenAI CTO Mira Murati, unveiled Inkling, its first internal AI model. Additionally, it is open-weight, which means that outside developers and businesses can download and alter it directly, in contrast to the flagship models from OpenAI, Anthropic, or Google.
Inkling is a mixture-of-experts system with 975 billion total parameters, yet it only uses roughly 41 billion of those for any given task. This typical design keeps very large models running more quickly and affordably. According to the company’s own release materials, it was trained on 45 trillion tokens of text, image, audio, and video, and reasoning natively across all four. However, its outputs are now restricted to text, which includes structured data, code, and stylized artifacts.
After a year and a half of developing AI infrastructure mostly behind closed doors, the model is Thinking Machines Labs’ first public demonstration. A May research preview of “interaction models”—AI created to listen and speak (and even disrupt) rather than stop and wait like normal chatbots—had already shown some of that work. It’s also a test of the startup’s main hypothesis, which is that AI that businesses can customize will perform better than the universally applicable models currently offered by the largest laboratories.
Inkling allows users to adjust “thinking effort” in order to sacrifice speed for measured solutions, such as indicating uncertainty instead of speculating. According to the business, in order to achieve the same coding performance on one benchmark, Inkling requires a third as many tokens as Nvidia’s Nemotron 3 Ultra, their most recent generation open-weight model.
Inkling is not the best in its class, according to Thinking Machines. Inkling is “not the strongest model available today, closed or open,” according to its briefing materials. Well-rounded performance and customization are clearly what it’s aiming for instead.
This begs the question of who this solution is truly intended for within the enterprise market. For the time being, Thinking Machines is presenting Inkling less as a final product and more as a starting point that businesses can refine using Tinker, the company’s model-customization platform. This implies that, for example, consumers are in charge of ensuring the security of their customizations rather than Thinking Machines. (Serious machine-earning talent is needed for fine-tuning.)
With ChatGPT, Claude, and Gemini, respectively, developed to compete as general-purpose chatbots first with agentic, autonomous capabilities overlaid on top, OpenAI, Anthropic, and Google have all adopted very different strategies.
This publication was obviously intended to be set against the backdrop of a piece that Thinking Machines published last week. The company claimed in that post that AI that is educated centrally by one company and then set in stone underperforms AI that companies shape themselves because so much of the expertise is unique to the individuals who possess it.
The argument is becoming more and more popular. Microsoft CEO Satya Nadella, whose company has invested billions in both OpenAI and Anthropic, cautioned in a blog post published on Sunday that businesses that use proprietary AI models essentially pay twice: once in subscription fees and again by providing business knowledge embedded in their prompts and corrections, which can be absorbed into future model versions.
In an interview with TechCrunch last week, Hugging Face CEO Clem Delangue made a similar forecast. According to him, the majority of production AI work will move to private or open-source alternatives—exactly the split Thinking Machines is built around—while frontier models will increasingly be saved for experimental and high-value activities.
A recent initiative with the largest hedge fund in the world, Bridgewater Associates, which isn’t, for the record, an investment in Thinking Machines, provided the most convincing evidence for the company’s claims. Researchers from both businesses used Bridgewater’s financial knowledge to further train an already-existing open-source model. Although those results are based on the two organizations’ internal review rather than an impartial one, the result was reported to score 84.7% on financial reasoning tests, outperforming top proprietary AI models while costing around a sixteenth as much to run.
In any case, Thinking Machines is highlighting the speed at which it arrived. It took Anthropic around three years and OpenAI about five to launch its technology and start making money. According to Thinking Machines, it completed the same task in roughly nine months.
Some will question whether Inkling was trained using outputs from rival models—a technique called “distillation” that has received criticism from the industry as a whole. According to the company’s own documents, the quick answer is partially. Pre-trained Thinking Machines: Before large-scale reinforcement learning took over, Inkling claims to have generated some of its early post-training data using other open-weight models, such as Moonshot AI’s Kimi K2.5. The business is adamant that fully self-contained post-training will be used in future models.
Thinking Machines has been more cautious when it comes to costs. It partnered with Nvidia in March to deploy a gigawatt of Vera Rubin processing capacity and trained Inkling exclusively on Nvidia’s GB300 NVL72 systems, but it hasn’t said how it intends to pay for those expenses, and by most accounts, income hasn’t been a top priority. (A $50 billion financing round was reportedly in the works last November, but by January it had stalled; the business has refused to discuss its funding situation since.)
A related concern is whether Thinking Machines will ever spend as much as OpenAI or Anthropic, or if its efficiency-driven strategy would make the economics different. In other words, unlike the metered access that OpenAI and Anthropic offer, once weights are made public, anyone who downloads them is not required to pay Thinking Machines to run them, so the company’s bet may be less that it will eventually spend like its bigger rivals than that it won’t need to at all. The company’s income must come from training, fine-tuning, and now a portion of the hosting ecology that surrounds Tinker, not from the model itself.
At least the headcount appears more stable. After a wave of departures earlier this year, including two co-founders who left for OpenAI in January, Thinking Machines now employs about 200 people.
For its part, Thinking Machines doesn’t appear to be as keen on showcasing specific actions as a large portion of the industry. An insider claims that the company’s culture intentionally prioritizes consistency above dependence on any one individual. It makes sense: if they were never elevated in the first place, switching teams is less of a setback. Given how much of its own history is still linked to the name of its now-famous co-founder, whether or not she intended it, it’s also a surprising thing for a firm to insist on.