Verified, then logged.
Sun 22 Nov 2026 · 00:22 KST / 15:22 UTC한국어 | ENSearch /

08 / Industry · Analysis

Inkling 2 Small weights carry distinct memory requirements

Inkling 2 Small's total and active parameter counts describe different scales, while its weight formats carry separate GPU memory requirements.

Rows of server racks, overhead power infrastructure, and cooling pipes line an empty aisle inside an unbranded AI data center. AI-generated image

Open Model Weights reported the release of Thinking Machines' Inkling 2 Small weights. It listed 594 billion total parameters. About 17 billion are active per token.

The report gave different memory requirements for each weight format. The Small name alone therefore does not establish the equipment needed. Model size and hardware requirements need separate readings.

Open Model Weights described the model as multimodal. It accepts text and images as inputs. Its output is text, so input and output support differ.

Total parameters and active parameters

Open Model Weights described Inkling 2 Small as a mixture-of-experts model. MoE models use selected parts of a model for each token. This distinguishes the full model from the portion active during processing.

In the report, total parameters means all parameters in the model. Active parameters means those used per token. The total count should therefore not be read as the count used for every token.

Weight formats and GPU memory

Open Model Weights listed NVFP4 and MXFP8 as the released checkpoint formats. Checkpoints are saved model weights used to load a model. Each format carries its own minimum GPU memory requirement.

The report attributed those minimum requirements to the model card. They refer to combined graphics processing unit (GPU) memory. Planning to run the model therefore requires separate decisions about obtaining the weights and securing suitable hardware.