Thinking Machines Lab Releases Open-Weights Multimodal Model Inkling
Mira Murati's Thinking Machines Lab has launched its first open-weights model, named Inkling. This multimodal model is a Mixture-of-Experts transformer with a total of 975 billion parameters, of which 41 billion are active. It is released under the Apache-2.0 license and was trained on 45 trillion tokens encompassing text, images, audio, and video.
A smaller version, Inkling-Small, with 276 billion parameters (12 billion active), is currently undergoing testing, with its weights expected to be released upon completion. The accompanying model card and training data documentation are notably brief, stating that the datasets include both public domain content and material subject to intellectual property protection, sourced from the open internet, public repositories, and third parties.
Thinking Machines Lab positions Inkling not as a cutting-edge model, but as a robust base model suitable for fine-tuning via their Tinker platform. Its strengths lie in its multimodal capabilities, efficient processing, and availability for customization. The release is seen as a significant addition to the US open-weights ecosystem, competing with models from China and joining other US contenders like NVIDIA Nemotron and Gemma 4. The model demonstrated its multimodal functionality by generating an SVG of a pelican on a bicycle and subsequently describing the generated image.
The release of Inkling by Thinking Machines Lab introduces a new contender in the open-weights multimodal model space, emphasizing customization and accessibility. By adopting the Apache-2.0 license, the lab promotes broader adoption and innovation, fostering an ecosystem where developers can build upon its capabilities. The strategy of offering a strong base model for fine-tuning on a proprietary platform like Tinker suggests a business model focused on enabling downstream applications rather than solely competing on raw model performance. This approach aligns with a growing trend of specialized AI development, where foundational models serve as platforms for diverse, tailored solutions. The limited transparency regarding training data, while common, highlights an ongoing tension between open-source principles and proprietary data sourcing in AI development, posing challenges for independent verification of model biases and capabilities.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.