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

Thinking Machines Lab has released its first foundation model, Inkling, with full open weights and developer fine-tuning through Tinker, SiliconANGLE reported.

The launch gives the AI startup a public model after a year in which its funding rounds and Nvidia partnership drew most of the attention.

Inkling is not being presented as a closed chatbot.

Developers can download, adjust and run the weights, while Tinker remains the paid service for fine-tuning open-weights models.

Inkling Uses 975 Billion Parameters And Open Weights

A company blog post describes Inkling as a mixture-of-experts model with 975 billion parameters.

Thinking Machines said an average prompt draws on about 41 billion parameters to process tasks faster and keep costs low.

The training description lists about 45 trillion tokens spanning text, image, audio and video.

Inkling can reason across all four inputs, but its outputs are limited to text, including code, styled artifacts and structured data.

Those full open weights let developers inspect and adapt the model code.

Thinking Machines also outlined thinking-effort controls for trading processing speed against accuracy, and SiliconANGLE wrote that the model flags uncertainty in outputs.

Mira Murati previously served as Chief Technology Officer of OpenAI before leaving in September 2024, according to the account.

Her stated focus on accessibility, customisation and multimodal collaboration appears in the launch, with public outputs still limited to text.

Tinker Carries The Fine-Tuning Revenue Model

Developers can fine-tune the model directly on Tinker, the startup's training API that launched in October, according to SiliconANGLE.

Rather than charging for metered access to the model itself, the paid API carries the revenue plan for the release.

The training path also extends the Nvidia connection.

Thinking Machines stated that Inkling was trained on Nvidia's GB300 NVL72 system under a partnership announced in March.

In early test results cited by the company, Inkling reached comparable coding performance with Nvidia's Nemotron 3 Ultra while using two-thirds fewer tokens, the startup claimed.

Independent benchmark methodology and customer deployment results are not included in that comparison.

Bridgewater Test Gives Inkling A Finance Example

SiliconANGLE cited a collaboration with Bridgewater Associates in which researchers used Tinker to fine-tune an open model with specialised financial data.

The resulting lightweight model scored 84.7% on financial reasoning benchmarks at less than 10% of the cost of advanced proprietary alternatives, the account said.

Futurum Group analyst Mitch Ashely told the Wall Street Journal, as cited in the account, that the open-weight model ecosystem had been dominated by Chinese AI firms for the last year.

The quoted assessment described the release as a Western alternative for enterprises weighing customisation economics and infrastructure control.

The lab acknowledged that its new model is not as strong as some advanced proprietary AI systems.

The release is positioned as a base model that organisations can fine-tune and run on their own infrastructure, not as a rigid chatbot application.

Thinking Machines developed the model from scratch in less than nine months, according to the account.

Thinking Machines did not provide independent benchmark validation for Inkling.