Thinking Machines Inkling Arrives: Mira Murati Open-Sources a 975-Billion-Parameter Model

by Zaira Varghese 16. July 2026 AI 0

Thinking Machines Inkling landed on Wednesday morning, and it is not the product anyone expected from a lab that has raised money at frontier-lab valuations. Mira Murati’s startup released its first in-house model as open weights under an Apache 2.0 license — downloadable, modifiable, deployable, with virtually no strings attached.

That is a deliberate swipe at the house style. OpenAI, Anthropic and Google keep their flagships behind APIs. Murati, who ran engineering at OpenAI as CTO before leaving to start Thinking Machines Lab, just gave hers away.

What’s inside Thinking Machines Inkling

The specs are substantial:

975 billion total parameters, in a mixture-of-experts architecture that activates only about 41 billion for any given task — the standard trick for keeping enormous models fast and cheap at inference time.
45 trillion tokens of training data spanning text, image, audio and video.
Native multimodal reasoning across all four modalities, per the company’s own release materials.
Apache 2.0, the permissive end of the licensing spectrum — no usage caps, no revenue thresholds, no acceptable-use rider with teeth.

On paper that puts it in the same weight class as Kimi and Nvidia’s Nemotron family, the two open-weight releases that have set the bar over the last year.

A starting point, not a finished product

The more interesting move is the positioning. Thinking Machines is marketing Inkling less as something you use and more as something you build on — raw material for organizations to fine-tune through Tinker, the company’s model-customization platform.

This is the whole thesis of the company, made concrete. Murati has argued since founding the lab that one-size-fits-all frontier models are the wrong shape for most real deployments, and that the value sits in adapting a strong base to a specific domain. Giving away the base and charging for the adaptation layer is a coherent way to bet on that. It also means the open release isn’t charity — it’s customer acquisition for Tinker.

Why open weights, why now

The competitive logic is hard to miss. Thinking Machines cannot out-distribute OpenAI or out-integrate Google. What it can do is be the model that shows up in every enterprise evaluation because there’s no procurement conversation to have first. Open weights are a distribution strategy for challengers, which is exactly why Meta reached for them and why the strongest Chinese labs have leaned on them.

VentureBeat’s coverage notes the model is also being pitched on low cost and “resistance to censorship” — a framing aimed squarely at buyers who don’t want their AI stack subject to another company’s content policy or pricing changes. That message plays well with regulated industries, sovereign customers and anyone who has watched an API provider deprecate a model mid-project.

What it doesn’t tell us yet

The release materials are the company’s own, and independent benchmarks haven’t caught up. Parameter counts and token volumes are inputs, not results, and the open-weight leaderboard has repeatedly punished models that looked dominant on a spec sheet. A 975B mixture-of-experts model is also not something a startup runs on a spare GPU — the practical audience for downloading it is narrower than “anyone.”

Whether Thinking Machines Inkling is genuinely competitive with the closed flagships, or merely competitive with other open releases, is a question the next few weeks of independent evaluation will answer.

The bottom line

This is the first real product from one of the most-watched and best-funded startups in AI, and it arrives as a gift with a business model attached. If Inkling holds up under scrutiny, Murati will have demonstrated that a lab can reach the frontier and still hand the weights over — which puts uncomfortable pressure on every competitor whose moat is the API in front of the model.

Read TechCrunch’s full report on the Inkling release for more on the launch.