Inkling

Inkling is a large Mixture-of-Experts transformer built for broad, flexible deployment, with 975B total parameters of which 41B are active per token. It was pretrained from scratch on 45 trillion tokens spanning text, images, audio, and video, giving it a genuinely multimodal foundation rather than a text-first core with bolted-on perception. The model supports a context window of up to one million tokens, allowing it to handle long documents, extended conversations, and large multi-file workflows without losing coherence. Thinking Machines Lab released the full weights publicly, positioning Inkling as an open base model that teams can fine-tune, distill, or adapt to specialized domains rather than a closed flagship competing on raw leaderboard scores.

A defining design choice is Inkling's variable thinking effort, which lets callers trade latency and cost against reasoning depth on a per-request basis, paired with native reasoning across images and audio alongside text. This makes it well suited to agentic tool-use pipelines, instruction-following tasks, and multimodal assistants where the same model must decide how hard to think depending on the query. The smaller sibling in the family, Inkling-Small, was trained with a comparable recipe at a fraction of the active parameters, so teams that need a lighter footprint can stay within the same architecture lineage. Practically, Inkling fits organizations that want a customizable, open multimodal foundation for long-context reasoning, multimodal understanding, and adaptive inference rather than those chasing the single strongest closed model.