Mira Murati spent a year and a half building in the dark. On July 15, 2026, Thinking Machines Lab finally turned on the lights — and the first thing it showed the world was Inkling, an open-weight model that quietly rejects almost everything the frontier labs stand for.
It is not the biggest model. It is not the smartest. Thinking Machines will tell you that itself. What makes Inkling interesting is the bet underneath it: that the future of enterprise AI belongs to models companies can own and reshape, not rent by the token.
What Inkling actually is
Inkling is a mixture-of-experts (MoE) system with 975 billion total parameters, but it activates only about 41 billion for any given task. That sparse design is now standard practice among large models — it keeps a huge network fast and cheap to run by lighting up a small slice of its weights per query.
The model was trained on 45 trillion tokens spanning text, images, audio, and video, and it reasons natively across all three modalities rather than bolting vision or speech on afterward. Two design choices stand out:
- Calibrated answers. Inkling is built to flag its own uncertainty instead of confidently guessing — a direct swipe at the hallucination problem that plagues chatbots.
- Adjustable thinking effort. Users can dial reasoning up or down, trading depth for speed when a task doesn't need heavy deliberation.
On one coding benchmark, Thinking Machines says Inkling reaches the same performance as Nvidia's Nemotron 3 Ultra while using roughly a third as many tokens — a claim about efficiency, not raw capability.
"Not the strongest model available today, closed or open." — Thinking Machines' own briefing materials
That is a striking thing for a lab to say about its own flagship. But it fits the strategy. Inkling isn't sold as a finished product. It's a starting point — something enterprises fine-tune for themselves through Tinker, the company's model-customization platform.
The bet against one-size-fits-all AI
The whole release is an argument. In a post published days earlier, Thinking Machines claimed that AI trained centrally by one lab and then frozen underperforms AI that organizations shape themselves, because so much valuable expertise lives inside specific teams and can't be captured by a generic model.
That argument is gaining powerful backers. Microsoft CEO Satya Nadella warned in a blog post that enterprises leaning on proprietary models effectively pay twice: once in subscription fees, and again by handing over the business knowledge embedded in thousands of prompts and corrections, which can be absorbed into the vendor's future model versions. Hugging Face CEO Clem Delangue made a similar prediction — frontier models reserved for experimentation and high-value work, while most production workloads shift to private or open alternatives.
The most concrete evidence came from a joint project with Bridgewater Associates, the world's largest hedge fund. Researchers took an existing open-source model, trained it further on Bridgewater's own financial expertise, and reported a score of 84.7% on financial reasoning tests — beating top proprietary models at roughly one-fourteenth the cost to run. Worth a caveat: those results come from the two companies' own evaluation, not an independent one.
The catches
An open-weight, customizable model sounds liberating. It also shifts real burdens onto the customer.
Fine-tuning Inkling through Tinker requires serious machine learning talent — this is not a consumer chatbot you open and start typing into. And because organizations do their own customization, they also own the safety of those customizations. When you fine-tune a model, you inherit responsibility for how it behaves.
There's also the distillation question hanging over the industry. Did Inkling learn from competitors' models? Partly. Thinking Machines pretrained Inkling from scratch but says it used other open-weight models — including Moonshot AI's Kimi K2.5 — to help generate some early post-training data before large-scale reinforcement learning took over. The company insists its next model will use fully self-contained post-training.
The business puzzle
Here's the part that should make you squint. Once model weights are public, nothing obligates the people who download them to pay Thinking Machines to run them. That's the opposite of the metered, per-token access OpenAI and Anthropic sell.
So where does the money come from? Tinker — training, fine-tuning, and a cut of the hosting ecosystem built around the model. It's a genuinely different business model, and it's unproven at scale.
The infrastructure bill is real. Inkling was trained entirely on Nvidia's GB300 NVL72 systems, part of a March partnership to deploy a gigawatt of Vera Rubin compute. Thinking Machines has been guarded about how that squares with revenue, which hasn't been its priority. The company does emphasize speed: it says it went from founding to a shipped, public model in about nine months, versus roughly five years for OpenAI and three for Anthropic. Headcount now sits around 200 people.
The Bottom Line
Inkling won't top the leaderboards, and Thinking Machines isn't pretending otherwise. Its value is as a thesis made real: that enterprises get more from a model they can own, adapt, and run cheaply than from a rented one that quietly learns from their data. Whether that thesis pays the bills is a separate question — one that rests on Tinker, not on Inkling itself. But for any organization with ML talent and data worth protecting, this is the first open-weight release in a while that's worth taking seriously as a foundation rather than a finished tool.
