Thinking Machines launches Inkling, its 975-billion-parameter open model

Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati, has unveiled Inkling, its first in-house artificial intelligence model. The launch is significant both for its scale and its approach: rather than selling a closed, general-purpose assistant, the company wants organizations to download the model, modify it and specialize it.

Inkling has 975 billion total parameters, the internal units a model adjusts during training. However, it activates only about 41 billion for each request. This architecture, known as mixture-of-experts, splits the work across specialized groups of parameters, allowing a very large model to avoid using its full capacity for every response.

An open model to adapt, not a finished chatbot

Inkling's weights are open: developers and companies can download them and fine-tune the model directly. That does not mean the entire process is open — the training data and infrastructure are not necessarily open — but it gives customers far more control than API access to models such as ChatGPT, Claude or Gemini.

The company trained Inkling on 45 trillion tokens from text, images, audio and video. A token is a unit of text processed by the system, similar to a word or part of a word. Although its training is multimodal, its outputs are currently limited to text: it can generate code, structured data and formatted documents.

Thinking Machines says the model is designed to express uncertainty rather than make up an answer when it lacks sufficient grounding. It also lets users adjust reasoning effort: they can request a faster answer or allocate more compute to a difficult task. This feature is increasingly common in advanced models, but is particularly useful when a company must balance cost, latency and accuracy across thousands of queries.

The company does not position Inkling as the most capable model on the market. Its own announcement acknowledges that it is not the strongest model available today, open or closed. The goal is different: to provide a sufficiently competitive foundation that an organization can tailor to its own needs.

Customization as the core product

That strategy relies on Tinker, Thinking Machines' platform for training and fine-tuning models. Fine-tuning involves continuing a model's training with a company's examples, documentation and criteria. A bank can teach it its internal language and analytical processes; an industrial company can provide its technical manuals and maintenance records.

The idea has a clear advantage: much of an organization's useful knowledge does not appear in the public data used to train large models. It resides in procedures, past decisions, proprietary terminology and exceptions accumulated over years. A general-purpose chatbot can help draft or summarize; a specialized model aims to take part in tasks where that context makes the difference.

But opening and customizing a model does not eliminate the challenges. It requires teams with machine-learning expertise, well-prepared data and in-house safety controls. If a company modifies the system's behavior, it also takes on much of the responsibility for assessing its failures, biases and potential misuse.

Efficiency and a test in finance

Thinking Machines says Inkling uses one-third as many tokens as Nvidia's Nemotron 3 Ultra to achieve the same performance on a coding benchmark. The comparison points to one of generative AI's biggest costs: it is not enough to train models; they must also be paid for every time a user runs them.

The company has also cited a project with Bridgewater Associates, the world's largest hedge fund. The two organizations fine-tuned an existing open model with Bridgewater's financial expertise and say it achieved 84.7% on financial reasoning tests, outperforming leading proprietary models at roughly one-fourteenth of the operating cost. The results were evaluated by the parties themselves, not by an independent lab, so they are more an indication of the strategy than a definitive industry comparison.

Inkling was trained entirely on Nvidia GB300 NVL72 systems. In March, Thinking Machines announced a partnership with Nvidia to deploy a gigawatt of Vera Rubin computing capacity, a sign that its efficiency proposition does not mean forgoing large-scale infrastructure.

The business is not charging for every query

Closed models typically turn every use into revenue: customers pay for a subscription or API access. An open-weight model changes that relationship. Once downloaded, it can run on a customer's infrastructure or through external providers without its creator charging for every response.

That is why Thinking Machines' business depends less on Inkling as a standalone product than on Tinker, training, fine-tuning and hosting built around the model. The company is betting that many organizations will not seek the most powerful general-purpose model, but one they can control, adapt and operate at predictable costs.

The question Inkling raises is not only who builds the best model. It is also who retains control over the knowledge that makes AI useful within a company. In that arena, open weights offer a concrete alternative to dependence on a small number of closed-model providers.