Inkling Small (Deep Infra)

Inkling-Small is the lighter sibling in Thinking Machines Lab's Inkling family of open-weights models, designed to extend the same multimodal foundation as the flagship Inkling while reducing cost and latency for practical deployment. It is a Mixture-of-Experts transformer that handles text, image, and audio inputs and produces text outputs, sharing the family's recipe of native multimodal reasoning trained across text, images, audio, and video. The smaller checkpoint is positioned as a flexible base for customization, reflecting the lab's mission of building models that "extend human will and judgment," and pairs naturally with the Tinker platform that the lab offers for fine-tuning and interactive experimentation.

Architecturally, Inkling-Small inherits Inkling's hybrid local-and-global attention design but in a more compact form, using 42 decoder layers with a hidden size of 4096, a 512-token sliding window for local attention, and every sixth layer applying global attention. Each sparse feed-forward layer routes tokens to 6 of 256 experts and additionally evaluates 2 shared experts, allowing efficient inference while preserving multimodal capability. This MoE structure, combined with the family's image, video, and audio encoders, gives Inkling-Small a balanced profile for teams that want a customizable open-weights model suited to agentic, reasoning, and tool-driven workflows without committing to the largest variant's footprint.