Thinking Machines Lab unveils 975B-parameter Inkling

Thinking Machines Lab has released its first general-purpose AI model, Inkling, and made it open for developers and organizations to download, run on their own infrastructure, and fine-tune for specific tasks.

The startup, founded by former OpenAI executives, says Inkling was not built primarily to top popular benchmarks. Instead, the company is positioning it as a flexible base model for customization. According to Thinking Machines Lab, its capabilities are comparable to the strongest open models currently coming out of Chinese companies, while its main advantage is adaptability.

Inkling uses a Mixture of Experts architecture. In total, the model has 975 billion parameters, but it activates only about 41 billion for any given task, cutting compute costs without a major hit to quality. It also supports a context window of up to 1 million tokens.

The model was trained from scratch to handle not just text, but also audio and video. The company also claims strong performance in reasoning and coding. For enterprise users, Inkling is available through the company’s Tinker platform, which is designed for fine-tuning models on corporate data. Alongside the flagship model, Thinking Machines Lab also announced Inkling-Small, a lighter version with 12 billion active parameters that is cheaper to run and faster in use.

Thinking Machines Lab said it also used Inkling itself during development for further fine-tuning and improvement, reflecting a broader pattern in which newer AI models are used to improve later versions of themselves.

Recommended reading

Hassabis says STEM makes you 10x better at AI

One unusual effect appeared during training: Inkling stopped producing intermediate reasoning in natural language on its own, apparently treating that style of explanation as an unnecessary computational expense. The developers later restored the capability so the model’s behavior would be easier for users to understand and inspect.

Thinking Machines Lab was founded in February 2025 by OpenAI alumni including former CTO and interim CEO Mira Murati, OpenAI co-founder John Schulman, who helped create ChatGPT, and former OpenAI vice president Lilian Weng, who led research in AI safety and robotics.

Last year, the startup raised a record $2 billion seed round at a $12 billion valuation. Before Inkling, the company had already introduced the Tinker model-customization platform, demonstrated a voice interaction system for AI, and published several machine learning research papers. Inkling is its first full-scale entry into the large language model market, built around the argument that AI systems should not remain under the control of a small number of developers but should be available for independent deployment, customization, and training on private data.