Inkling (Deep Infra)

Inkling is Thinking Machines Lab's first model trained from scratch, designed as a broad foundation that downstream developers can fine-tune and adapt rather than a single benchmark-leading system. It is built as a Mixture-of-Experts transformer with 975B total parameters and 41B active per token, pretrained on roughly 45 trillion tokens spanning text, images, audio, and video. The model reasons natively over text, images, and audio, supports controllable thinking effort, and accepts a multimodal input stream while emitting text. A lighter preview sibling, Inkling-Small, was released in parallel at 276B total parameters with 12B active, sharing the recipe for lower cost and latency use cases.

Practically, Inkling is aimed at developers building agentic and tool-use applications, coding assistants, chatbots, and retrieval-augmented generation systems, and the open-weights release under Apache 2.0 invites third-party customization, including fine-tuning on the Tinker platform with full weights hosted on Hugging Face. Its long context window supports extended document and conversational workflows, and the modality coverage plus reasoning controls make it a flexible base for instruction-following and general conversational tasks across English and other languages. For teams that want a customizable multimodal foundation rather than a turnkey closed model, Inkling fits as a starting point that can be specialized to a target product or domain.