Thinking Machines Lab has released Inkling 2 Mini, a 26.4-billion-parameter open-weights transformer designed to cut inference costs through an unconventional architectural choice: looping its middle layers three times per token during processing.
The design targets a core pain point in LLM economics. As AI workloads scale, inference costs—driven by token consumption—become a material line item for enterprises. Inkling 2 Mini aims to deliver equivalent reasoning performance while reducing total tokens processed, directly lowering per-inference expense and latency.
As an open-weights model, Inkling 2 Mini can be downloaded, inspected, and fine-tuned by enterprises, in contrast to closed proprietary models that lock users into API-based access and recurring licensing fees. This architectural choice enables on-premise or private cloud deployment, giving companies direct control over data residency and inference spend without cloud provider markup.
The 26.4-billion-parameter scale positions the model for enterprise tasks including customer service automation and data analysis workflows where reasoning quality matters but speed and cost have become hard constraints.

