When you buy through affiliate links in our content, we may earn a commission at no extra cost to you. Learn 
Thinking Machines Lab has introduced Inkling, the startup’s first AI model—and, perhaps more importantly, the first concrete proof of what former OpenAI CTO Mira Murati has been building since leaving the tech firm.
For months, the startup had been the AI world’s most intriguing mystery box. It teased lofty ideas about “human judgment,” showed off futuristic AI interactions, and published research papers that read like a manifesto. But there was one obvious omission.
An actual AI model.
That changed this week.
On Wednesday (July 15), the company unveiled its first open-weights foundation model — a multimodal Mixture-of-Experts system with a whopping 975 billion total parameters (41 billion active), a 1 million-token context window, and training spanning 45 trillion tokens across text, images, audio and video.
The specs are eye-catching. The philosophy behind them is arguably even more interesting.
Thinking Machines Lab introduces anti-‘one model rules them all’ pitch with Inkling
The company is apparently trying to sell Inkling as the smartest AI on Earth.
Inkling is not the strongest overall model available today, open or closed. Instead, a combination of qualities makes it a good open-weights base for customization.
It explicitly acknowledges that Inkling is “not the strongest overall model available today,” but positions it as a highly capable foundation model that’s designed to be customized rather than treated as a finished product.
It’s a subtle but notable change in an industry obsessed with benchmark bragging rights.
Instead of chasing the latest leaderboard crown, Murati’s team is betting that developers increasingly want models they can actually reshape for their own needs. They state: “We want to make customization accessible for more use cases… releasing a model we trained from scratch with the full weights available, so that people can make it their own.”
It’s a philosophy the company has been hammering home for months. Last week (July 10), Thinking Machines argued that today’s AI is trained in a few centralized locations and then essentially frozen, while future AI should evolve alongside the people using it. The goal, it says, is AI that “extends human will and judgment” rather than replacing it.
Translation: don’t just rent intelligence—own it, tweak it and make it weird.
Open weights, but make it practical
Inkling is being released with open weights and integrated into Thinking Machines’ own fine-tuning platform, Tinker, where developers can customize the model without starting from scratch.
A model that’s confident in every answer it gives, including when it’s missing info and confabulates, forces the user to double-check everything.
The company even couldn’t resist a little AI theatre.
To demonstrate the system, Inkling reportedly fine-tuned itself. Using Tinker, the model wrote its own fine-tuning job, generated synthetic training data, ran the training process and evaluated the results. It’s either an impressive showcase or the beginning of AI’s annual performance review cycle—possibly both.
Built for collaboration and automation
And if you’ve followed Thinking Machines since launch, Inkling feels less like a surprise than the missing puzzle piece.
In May, the company introduced “interaction models,” arguing that AI should behave less like an email correspondent and more like a colleague sitting across the table.
These systems are said to be designed for continuous, real-time interaction across voice, video and text, complete with interruptions, simultaneous conversation and asynchronous reasoning happening in the background.
The company says one of Inkling’s primary design goals is to serve as the background reasoning model powering those interaction systems, combining deep reasoning with multimodal understanding.
More marathon than sprint
The system is a multimodal model from the outset, handling text, images and audio natively. It also supports what Thinking Machines calls “controllable thinking effort,” allowing developers to trade off reasoning depth against latency and token costs depending on the task.
The release highlights agentic coding, browser automation, document generation and multimodal reasoning, while benchmark charts suggest competitive performance across coding, reasoning, vision and audio tasks without claiming outright dominance.
Thinking Machines is essentially saying: Here’s a really capable foundation. Now go make it yours.
NVIDIA’s Principal research scientist, Leon Derczynski, called it a “solid US-origin open model.” However, others were a little more skeptical.
Whether developers buy into the vision of open weights, extensive customization, and humans kept firmly in the loop remains to be seen.
But after months of philosophical essays and ambitious research previews, Thinking Machines appears to have something far more tangible to point at.
A model with a name, some very large numbers attached to it—and an invitation for everyone else to start tinkering.
