source: simon willison: inkling: our open-weights model
level: technical
thinking machines lab, led by mira murati, released inkling, their first open-weights model. it is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active parameters. the model is multimodal, trained on 45 trillion tokens covering text, images, audio, and video. it is licensed under apache 2.0. a smaller version, inkling-small with 276 billion total and 12 billion active parameters, is still being tested and will be released later.
the model card and training data documentation are notably brief. the lab states that training data includes public domain content and publicly available internet data, along with some third-party datasets. they acknowledge that some data may be subject to intellectual property protections. the lab does not claim inkling is a frontier model. instead, they position it as a strong base for fine-tuning using their tinker training platform, emphasizing its multimodal capabilities and efficient thinking.
the release adds a new us-based open-weights contender alongside models like nvidia nemotron and google gemma 4. early tests show inkling can generate simple svg images and describe them, though with some inaccuracies, such as misidentifying a pelican as a stork. the model is accessible via an api, and the open license allows broad use and customization.
why it matters: it provides a permissively licensed, large-scale multimodal base model for fine-tuning, expanding options for ai developers and researchers outside of chinese open-weight models.