Thinking Machines releases Inkling, its first open-weight AI model

Thinking Machines Lab released Inkling, its first proprietary AI model, giving investors another signal that the AI market is splitting between closed subscription products and customizable models companies can run and adapt themselves. The startup, founded by former OpenAI CTO Mira Murati, is betting that enterprises will pay for tools around customization rather than rely only on general-purpose chatbots.

The company said Inkling is open-weight, meaning outside developers and businesses can download the model’s weights, the internal numerical settings that shape how an AI system behaves, and modify them directly. That differs from the flagship models behind OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini, which are sold mainly through controlled access rather than direct model downloads.

According to Thinking Machines, Inkling uses a mixture-of-experts design with 975 billion total parameters. Parameters are the values a model learns during training. In this setup, the model activates about 41 billion parameters for a given task, which the company says helps keep a very large system more efficient to run.

Thinking Machines said Inkling was trained on 45 trillion tokens across text, image, audio and video data. Tokens are the chunks of information AI systems process. The company describes the model as multimodal, meaning it can work across different types of media rather than only text.

A customization-first pitch

Inkling is the company’s first major public product after roughly a year and a half of building mostly outside public view. Thinking Machines previously showed a May research preview of “interaction models,” which it described as AI systems designed for more natural spoken exchanges, including interruptions, instead of the turn-by-turn structure common in chatbots.

The company is positioning Inkling as a starting point for organizations using Tinker, its model-customization platform. Fine-tuning, or further training a model on specialized data, can help a company adapt AI to its own work. It also means customers need machine learning expertise and must manage safety issues tied to their own changes.

Thinking Machines said Inkling is designed to give calibrated responses, including signaling uncertainty rather than guessing. The company also said users can adjust “thinking effort,” a setting meant to trade off speed and reasoning depth. On one coding benchmark, Thinking Machines said Inkling used one-third as many tokens as Nvidia’s Nemotron 3 Ultra to reach the same performance.

The company did not claim Inkling leads the market. In its materials, Thinking Machines said the model is “not the strongest model available today, closed or open,” and framed it instead as a broad, adaptable system.

The enterprise AI cost argument

Thinking Machines’ launch lands as more AI executives question whether companies should keep renting closed systems. In a recent blog post, Microsoft CEO Satya Nadella said enterprises using proprietary AI models can pay twice: once through subscriptions, and again by contributing business knowledge through prompts and corrections that may improve later model versions.

Hugging Face CEO Clem Delangue made a similar point in comments to TechCrunch, saying frontier models may be used more for experiments and high-value tasks, while production work shifts toward private or open-source alternatives.

Thinking Machines has also pointed to work with Bridgewater Associates, the hedge fund, as evidence for its approach. The two companies said researchers trained an existing open-source model further on Bridgewater’s financial expertise, producing a model that scored 84.7% on financial reasoning tests and cost about one-fourteenth as much to run as top proprietary systems. Those results came from the companies’ own evaluation, not an independent review.

On training, Thinking Machines said it built Inkling from scratch during pretraining but used other open-weight models, including Moonshot AI’s Kimi K2.5, to generate some early post-training data before large-scale reinforcement learning. Reinforcement learning is a training method that rewards a model for preferred outputs. The company said its next model will use self-contained post-training.

Thinking Machines has disclosed a March strategic partnership with Nvidia to deploy a gigawatt of Vera Rubin computing capacity and said Inkling was trained on Nvidia GB300 NVL72 systems. Nvidia said it made a “significant investment” in the company as part of that partnership. Thinking Machines now has roughly 200 employees, according to TechCrunch.

This story draws on original reporting from TechCrunch.