Thinking Machines Debuts Inkling: A Giant Open Model
Mira Murati’s lab has released Thinking Machines Inkling, its first model. It is open-weight, enormous, and, by the company’s own admission, not the best one out there.
July 15, 2026 – 8:17 pm
Image by: Patrick T. Fallon
"We believe in keeping the weirdness alive." This line comes from a manifesto Mira Murati’s lab published last week. It’s also the thinking behind the lab’s first model, Inkling.
Thinking Machines Lab, founded by the former OpenAI chief technology officer, has released Inkling. It is open-weight, so any developer or company can download the model and reshape it—a stark contrast to the flagships sold by OpenAI, Anthropic, and Google.
Inkling is massive: a mixture-of-experts system with 975 billion total parameters, using only about 41 billion for any given task. It handles up to a 1 million token context window and trained on 45 trillion tokens of text, images, audio, and video. It reasons across text, images, and audio—but for now, it only writes text back, including code and structured data.
A Model That Admits It’s Not the Best
Here’s the twist: Thinking Machines does not claim Inkling is the best model available today, closed or open.
The lab is chasing range and adaptability, aiming to provide a broad, balanced base that organizations can fine-tune for their work, not a finished chatbot. Users can adjust its “thinking effort” to balance accuracy and speed. In one coding test, Inkling matched Nvidia’s Nemotron 3 Ultra using a third as many tokens.
The lab also previewed Inkling-Small, a lighter model with 12 billion active parameters, suitable for tasks where cost and speed are paramount.
The Bet: Shape It Yourself
The whole release hinges on one bet: AI trained in one place, frozen, loses to AI each organization can shape around its own expertise. Customers fine-tune Inkling through Tinker, Thinking Machines’ customization platform, and they own the result—along with any safety risks that come with it.
The lab points to a project with hedge fund Bridgewater as proof. They trained an open model on Bridgewater’s financial know-how, achieving 84.7% on financial reasoning tests—outperforming top proprietary models at a fraction of the cost. This figure comes directly from the two companies’ evaluation, not an independent one.
This argument is gaining traction. Microsoft’s Satya Nadella recently warned that firms using closed models pay twice, once in fees and again by handing over their knowledge in prompts. Cheap open-weight models from China are also pulling this direction.
Nine Months, With Some Borrowed Help
Thinking Machines emphasizes its speed. OpenAI took about five years to ship and earn, Anthropic roughly three, according to TechCrunch. Murati’s lab says it did it in about nine months.
To kickstart Inkling’s training, the lab leaned on other open models, including Moonshot’s Kimi K2.5.