Inkling is a general-purpose multimodal model from Thinking Machines Lab, built as a 66-layer sparse Mixture-of-Experts transformer with 975B total parameters and roughly 41B active at inference. That mixture design lets the model route each prompt to a focused subset of experts, supporting deep reasoning over text, image, and audio inputs while keeping generation cost closer to that of a much smaller active network. The Apache 2.0 release of the weights means developers can self-host, fine-tune, or audit the model, which makes it attractive for teams that want transparency about how their assistant reasons, especially for code generation and tool manipulation tasks where the model is positioned as a practical, adaptive reasoning system.
The 1M-token context window is the most distinctive practical feature, allowing Inkling to ingest very long codebases, multi-document analyses, or extended multi-turn sessions without losing earlier material. It pairs this long-context capability with multimodal input so the same model can ground answers in images, diagrams, or audio alongside text, and it is exposed through providers that offer private, zero-retention hosting as well as the standard API path. For builders, the fit is clear: long-running agents that need to remember and reason across large files, multimodal assistants that combine visual and textual information, and open-weights projects that want a strong generalist base to fine-tune for specialized domains.
