A former OpenAI CTO has shipped a frontier-class model fully open. Thinking Machines Lab, founded by Mira Murati, released its first model, Inkling, on July 16, 2026, publishing the weights of a 975-billion-parameter Mixture-of-Experts model under the Apache 2.0 license. It was trained on 45 trillion tokens, supports a 1 million token context, and can be downloaded immediately from Hugging Face. ASAP summarizes the announcement from primary reporting.

Murati took a different road from Altman

Thinking Machines Lab released its first product, Inkling, as fully open weights on July 16, 2026. Mira Murati, who was OpenAI's chief technology officer before leaving to found Thinking Machines Lab, shipped that first model not as a closed API but as open weights under the Apache 2.0 license, which is the heart of the announcement. It cuts directly against OpenAI's line of keeping its top models weight-closed. The weights are on Hugging Face and run out of the box on open-source inference stacks like vLLM, SGLang, and Llama.cpp. The company says training used NVIDIA GB300 NVL72 systems.

Inkling's design and its "self-repairing" trait

Inkling's architecture is built on two axes, scale and a self-improvement capability. It is a Mixture-of-Experts with 975 billion total parameters that activates 6 of its 256 routed experts plus 2 shared experts per token, for about 41 billion active parameters. The training data spans 45 trillion tokens across text, images, audio, and video, and the context window reaches 1 million tokens. Thinking Machines Lab says the model "is capable of writing its own fine tuning scripts to refine its behavior, teach itself new skills, and evaluate its abilities." The companion model released alongside it, Inkling-Small, has 276 billion total and 12 billion active parameters, built for lighter deployment.

The unusual admission that it isn't the strongest

The most striking part of this launch is not a spec but a stance. Thinking Machines Lab flatly stated that Inkling is "not the strongest overall model available today, open or closed." It is rare for a product launch to open by saying its own model is not the best. That admission signals two things. First, it is a positioning choice to compete on openness and customizability rather than on a benchmark crown. Second, avoiding an inflated top-tier claim actually raises the credibility of the numbers it does publish. Declining to call itself the strongest is, in itself, this team's differentiating message.

Why Apache 2.0 is the real bet

A single line of licensing decides the weight of this launch. Inkling ships under Apache 2.0, which is a different animal from the "modified licenses" or usage-capped community licenses that recent open models often attach. Apache 2.0 is a genuinely open license that clearly permits commercial use, redistribution, and patent grants, so a company can drop it into an internal product without a legal review marathon. Releasing a frontier-class model under these terms is uncommon. Even without a world-best score, "you can take full control with legal peace of mind" is a value a closed API cannot offer. Murati's team wagered on trust and freedom instead of raw performance.

A business designed around fine-tuning

Inkling did not arrive as a standalone model but bundled with a fine-tuning platform. Thinking Machines Lab designed the model to be customized through its Tinker platform and made it accessible through third-party services like TogetherAI, Fireworks, Modal, Databricks, and Baseten. One reported case notes that Bridgewater Associates fine-tuned an open model to its own data through Tinker and reached 84.7% on financial reasoning tests. The strategy this structure implies is clear. Instead of selling "the smartest general model," it sells a base model that companies can easily tame with their own data, plus the tools to do it. The model's trait of writing its own fine-tuning scripts fits the same fine-tune-first line.

What Korean teams should try now

For Korean organizations, Inkling is an especially practical option under an Apache 2.0 license. Because it is Apache 2.0, self-hosting on in-house GPUs keeps data from leaving the building while allowing free commercial use, with low license risk and a lighter legal-review burden. With about 41 billion active parameters, its inference load is relatively light for a 975-billion-parameter model, and Inkling-Small is there when a lighter deployment is needed. The key, given that general performance is not top-tier, is to measure directly on representative tasks whether fine-tuning on your own domain data reaches a usable level for real work. The benefit of openness becomes an asset only after that hands-on measurement.

Open questions that remain

Inkling's true ability is still an open question that awaits independent verification. Since Thinking Machines Lab said it is not the strongest, its exact benchmark position will emerge through community reproduction, and Bridgewater's 84.7% came from a specific fine-tuning setup, so it is early to generalize. How reliably the self-fine-tuning-script trait works in practice also needs verification. Even so, simply by opening a frontier-class scale under Apache 2.0, Inkling has widened the landscape of alternatives to closed models. The next step is community fine-tuning and re-measurement by a neutral harness.

Source: Thinking Machines Lab Inkling release (July 16, 2026; 975 billion total MoE parameters, 6 of 256 routed experts plus 2 shared active per token, about 41 billion active parameters, 45 trillion training tokens, 1 million token context, Apache 2.0 license, Inkling-Small at 276 billion total and 12 billion active, trained on NVIDIA GB300 NVL72, distributed on Hugging Face), based on the announcement and reporting by The Register and aiangst, summarized by ASAP.