Thinking Machines Lab, led by Mira Murati, has published its first open‑weights model, called Inkling. The lab describes Inkling as a Mixture‑of‑Experts transformer with a total of 975 billion parameters, of which 41 billion are “active.” The model was trained on multimodal data — roughly 45 trillion tokens of text, images, audio and video.
Purpose and intended use
Thinking Machines Lab says Inkling is not intended to be a top‑performing frontier model but rather a robust base model for customization and fine‑tuning. The lab recommends using its Tinker platform for downstream fine‑tuning. Key attributes highlighted by the team are the model’s multimodal capabilities, computational efficiency, and the fact that the weights are available under an Apache‑2.0 license.
Sizes and upcoming variant
In addition to the released large model, the lab is developing an Inkling‑Small variant. According to the announcement, Inkling‑Small will have 276 billion parameters with 12 billion active parameters. That smaller model is still under testing; the weights will be published once testing is complete.
Documentation and training data
The model card provided with Inkling is shorter than the documentation typically published by major U.S. AI labs, and the linked Training Data Documentation is also concise. The documentation states that the datasets used include public‑domain content as well as material that may be subject to intellectual property protection. Thinking Machines Lab says the services were developed using publicly available content from the open internet and public data repositories, and that certain datasets were acquired from third parties.
License and ecosystem context
Inkling is released under the Apache‑2.0 license, a notable detail for the open‑weights ecosystem in the United States. Thinking Machines Lab positions the model as a competitive open‑weights offering alongside other projects such as NVIDIA Nemotron and Gemma 4.
Practical examples and multimodal behavior
The lab demonstrated practical uses via its API: the model can generate SVG artwork and can describe uploaded images in detail. In one example, a curl request to the Thinking Machines API asked the model to generate an SVG of a pelican (or pelican‑like bird) riding a bicycle; in another step the produced image was rendered to JPEG and the model provided a detailed description, characterizing the bird as a “stork or seagull” and describing the scene in plain, colorful vector‑style terms.
Summary
Inkling represents a significant release: a large, multimodal, open‑licensed base model aimed at fine‑tuning and customization via the Tinker platform. The lab does not present it as a frontier‑level model, but its open weights and Apache‑2.0 license make it a useful resource for researchers and organizations looking for a modifiable multimodal foundation. The smaller Inkling‑Small will follow once testing concludes.
