What Is Thinking Machines Inkling? Mira Murati's First AI Model

Built to Be Owned, Not Just Rented

Inkling is the first foundation model from Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati in early 2025. Released July 15, 2026, it's a mixture-of-experts transformer with 975 billion total parameters and roughly 41 billion active per token, trained from scratch on about 45 trillion tokens spanning text, images, audio, and video. It reasons natively across those input types, though its outputs are currently limited to text — including code, styled artifacts, and structured data.

The full weights are open and available on Hugging Face, live for fine-tuning today on Tinker, the company's own customization platform. Thinking Machines is also previewing Inkling-Small, a lighter-weight sibling whose weights will follow after further testing.

Customization Over the Leaderboard

Thinking Machines is explicit that this is a different bet than most labs are making: that enterprises increasingly care less about renting the single smartest general-purpose model and more about owning one they can shape around their own data, infrastructure, and cost constraints — without needing to see or share what other customers are doing with the same base model.

A Western Answer to the Open-Weight Gap

Inkling lands at a moment when open-weight AI has been dominated almost entirely by Chinese labs.

Open-weight AI in 2026 has been overwhelmingly dominated by Chinese labs — DeepSeek, Qwen, GLM, and others — while Meta has scaled back its own open Llama ambitions in favor of a more proprietary approach. Inkling is one of the more credible Western attempts to fill that gap, giving companies wary of routing sensitive workloads through Chinese-hosted models, or through a handful of closed U.S. labs, a genuine third option.

Why "Owning the Model" Actually Matters to Some Companies

The distinction between "owning" a model and "renting access" to one via API sounds abstract until a company hits a specific real-world constraint that only ownership solves. If a business operates in a regulated industry — healthcare, finance, defense-adjacent work — sending customer data to a third-party API, even a well-secured one, can violate data residency and compliance requirements that have nothing to do with how good the model itself is. Owning the weights means the model can be deployed entirely within a company's own infrastructure, with data never leaving systems the company directly controls, which is a compliance and legal answer as much as a technical one.

The second practical reason ownership matters is customization depth. An API-accessed closed model can be fine-tuned or prompted in limited ways, but a company that owns the underlying weights can retrain, modify, and specialize the model far more deeply — adapting it to proprietary internal terminology, unusual data formats, or a specific workflow that a general-purpose API model was never designed around. For companies whose competitive advantage depends partly on how well their AI understands their specific business, that deeper customization can be worth the additional infrastructure investment required to actually run the model themselves.

The trade-off, of course, is that owning a model means owning the operational burden that comes with it — hosting, scaling, updating, and securing the infrastructure yourself, rather than letting a vendor handle all of that behind an API. For smaller companies without dedicated ML infrastructure teams, that overhead can easily outweigh the benefits of ownership, which is why Inkling's target market skews toward larger, more technically resourced organizations rather than small businesses or individual developers.

That target audience also explains the emphasis on Tinker, the company's fine-tuning toolkit, as a first-class part of the product rather than an afterthought. A model without a solid path to customization is only half-useful to the kind of organization Inkling is aimed at — the ability to actually adapt the model to a specific domain is often the entire point of choosing ownership over API access in the first place.

Who Inkling Is For

Frequently Asked Questions