Thinking Machines Lab, founded by Mira Mulati, former Chief Technical Officer of OpenAI, has released the first multi-modular model, Inkling. The model has been completed from zero training and the full weight has been made public in Hugging Face, using Apache 2.0 licences, and has been accessed on the company cloud platform Tinker for fine-tuning by developers.

Model size close to trillions of parameters

Inkling uses a hybrid expert structure, with a total parameter volume of 97.5 billion, which activates approximately 41 billion parameters per mission. Models support text, image and audio input, with a context window of 1 million token and pre-trained data of 45 trillion token, covering text, images, audio and video.

This is also the first model product that was launched after Mullati left OpenAI. She left in September 2024 and founded Thinking Machines Lab in February 2025. The company completed $2 billion in financing in July 2025, valued at $12 billion, and subsequently sought to continue financing on a $50 billion basis, but negotiations broke down in early 2026.

Agent's performance is ahead of Western class.

From the results of the tests disclosed, the advantages of Inkling are concentrated on proxy tasks. In MCP Atlas, the Inkling score was 74.1 per cent, nearly 30 percentage points higher than the Nvidia Nemotron 3 Ultra; in SWE-Bench Verified, the Inkling score was 77.6 per cent, or 70.7 per cent higher.

  • MCP Atlas score 74.1%
  • SWE-Bench Verified score 77.6%
  • All weights are open for download

MCP Atlas, which measures the stability of the primary measurement model through external tools, is used to test whether the model is capable of autonomously repairing real software deficiencies in GitHub. Thinking Machines Lab locates Inkling as a generic model rather than a single task optimization.

The Chinese model is still in the lead on some of the benchmarks.

However, Inkling is not the strongest model of the current overall performance. By comparison, Z.ai's GLM 5.2 scored 82.7 per cent on Terminal Bench 2.1, up from 63.8 per cent in Inkling; Kimi K2.6 retained its lead in such difficult scientific reasoning tests as Humanity's Last Exam.

In the safety test FORTRESS Adversarial, the Inkling score was 78.0 per cent, one of the best open source weight models in comparison with the sample. Thinking Machines Lab also predicted a smaller version of Inkling-Small, with a total parameter of 276 billion and a activated parameter of 12 billion. According to the company, the version was close to large model performance in most of the reasoning tests, but the time for the publication of weights had not yet been published.