Heatmap Analysis Exposes Attention Splitting in Inkling 2 Small

A community analysis of Thinking Machines' Inkling 2 Small visualized all 112 learned depth-wise queries via heatmaps, finding significant differences in split q/k/v stream queries in later layers, offering new insight into the model's attention structure.

2026-11-13 ~ 2026-11-13 · 2 related posts