Thinking Machines releases open-weight Inkling model

Thinking Machines Lab Inc., the AI startup founded by former OpenAI CTO Mira Murati, released Inkling, its first foundation model trained from scratch, with full open weights available for developers to download and fine-tune. The company is not charging for metered model access in the way closed API providers do, and instead plans to make money from Tinker, its paid fine-tuning service, though pricing was not disclosed.

Inkling is a mixture-of-experts model with 975 billion parameters, according to Thinking Machines. For a typical prompt, the company said the model activates about 41 billion parameters, a design intended to reduce compute cost and latency compared with using the full model on every request.

The company said Inkling was trained on about 45 trillion tokens spanning text, images, audio and video. It can process those input types natively, according to Thinking Machines, but its outputs are text-only, including code, structured data and styled artifacts.

The release puts Thinking Machines into the open-weight model fight at a time when Western options have thinned. Chinese AI companies have been prominent in open-weight releases over the past year, while Meta Platforms has moved away from the more open posture associated with earlier Llama models. Thinking Machines is positioning Inkling as an enterprise-controlled base model rather than a chatbot app.

The model includes controls for “thinking effort,” which the company said let developers trade speed for accuracy. Thinking Machines also said Inkling can mark uncertainty in its responses, a feature aimed at reducing confident wrong answers, though the company acknowledged the model does not match the strongest proprietary systems.

Developers can fine-tune Inkling through Tinker, the training API Thinking Machines launched in October. In early testing cited by the company, Inkling delivered coding performance comparable to Nvidia’s Nemotron 3 Ultra while using two-thirds fewer tokens. The company did not provide broader independent benchmark results in the announcement.

Thinking Machines said it built Inkling in less than nine months. The model was trained on Nvidia’s GB300 NVL72 system under the companies’ partnership announced in March. Nvidia previously made an investment in Thinking Machines, though the amount was not disclosed in the material cited by the company.

Murati’s startup has drawn attention less for products than for capital. Thinking Machines raised a $2 billion seed round in 2025 and has promoted accessibility, customization and multimodal collaboration as core themes. Inkling is the first major test of whether that positioning translates into developer adoption.

Futurum Group analyst Mitch Ashley told The Wall Street Journal that Inkling gives Western enterprises an alternative to Chinese open-weight models and could shift spending from per-token APIs to infrastructure controlled by customers. He also argued that choosing a base model becomes an architecture decision because fine-tuning creates switching costs over time.

The company pointed to work with Bridgewater Associates as evidence for its fine-tuning pitch. Researchers used Tinker to adapt an open model with specialized financial data, producing a lightweight model that Thinking Machines said scored 84.7% on financial reasoning benchmarks and cost less than 10% of proprietary alternatives.

Constellation Research analyst Holger Mueller told SiliconANGLE that the business model may be more notable than the model itself. His view was that charging for customization tooling, rather than raw model access, could push large language models further toward commodity infrastructure if customers adopt the approach.

This story draws on original reporting from SiliconANGLE.