Thinking Machines Releases Inkling Open-Weights Model With 975 Billion Parameters

Thinking Machines Lab has released Inkling, its first foundation model trained entirely from scratch, with full open weights that developers can download, inspect, fine-tune and run on their own infrastructure.

The model is available for customization through Tinker, the company’s paid training application programming interface, SiliconANGLE reported.

The release gives Mira Murati’s AI startup a public model after a year in which its sizable funding rounds and partnership with Nvidia Corp. drew most of the attention.

It also puts a Western open-weight system in front of companies looking for alternatives to lower-cost Chinese AI models.

What developers can use

Thinking Machines described Inkling as a mixture-of-experts model with 975 billion parameters.

An average prompt uses about 41 billion of them, the company said, allowing the system to process tasks faster and keep costs lower.

The model was trained on about 45 trillion tokens covering text, image, audio and video.

It can reason natively across all four input types, although its outputs are limited to text.

Those outputs can include code, styled artifacts and structured data.

Because the full weights are available, developers can inspect the model and adjust it for different use cases without paying expensive licensing fees.

Inkling also includes “thinking effort” controls that let developers trade processing speed for accuracy.

The model can flag uncertainty in its outputs instead of simply presenting uncertain responses as fact.

Murati, who was OpenAI’s chief technology officer before leaving in September 2024, has described her new company’s focus as accessibility, customization and multimodal collaboration.

Inkling’s open-weight release follows that approach, while keeping the user-facing result text-only.

How deployment works

Developers can fine-tune Inkling directly on Tinker, which launched in October.

Thinking Machines plans to generate revenue through that paid service rather than by charging customers for metered access to the model through an API.

The model also extends the company’s Nvidia relationship.

Thinking Machines said Inkling was trained on Nvidia’s GB300 NVL72 system under a partnership announced in March.

The company’s early tests showed comparable coding performance with Nvidia’s Nemotron 3 Ultra while using two-thirds fewer tokens.

Those results were presented by Thinking Machines, and the source did not include independent benchmark methodology or customer deployment results for the comparison.

Early tests and commercial boundary

A collaboration with Bridgewater Associates offers a separate example of the training model.

Researchers used Tinker to fine-tune an open model with specialized financial data, producing a lightweight system that scored 84.7% on leading financial reasoning benchmarks at less than 10% of the cost of advanced proprietary alternatives.

Futurum Group analyst Mitch Ashely told the Wall Street Journal that Chinese AI firms had dominated the open-weight model ecosystem for the previous year.

He described Inkling as a Western alternative for enterprises that want to shift spending from per-token API pricing to infrastructure they control, while treating the model they fine-tune as part of their software architecture.

Thinking Machines acknowledged that Inkling is not as strong as some of the most advanced proprietary AI systems.

The company is positioning it instead as a base model that organizations can customize and operate themselves, with Tinker providing the paid path for that fine-tuning.

Thinking Machines said it developed Inkling from scratch in less than nine months.