Strata Now Shows Which AI Model Experts Most Frequently Use Video Memory

Strata now includes an Experts tab that visualizes the operation of experts in the Qwen3.8-Flash-Next model. The article’s author integrated it with an existing profile file: the map displays layers and experts, with a frame highlighting nodes currently in video memory. No changes to the engine itself were required.

Strata runs the Mixture-of-Experts (MoE) model on a graphics card with limited memory: the most in-demand experts remain in VRAM, while others are loaded from RAM or SSD. Initially, the map colored cells based on current rank rather than absolute access frequency. Due to a decaying counter, rarely used nodes could shift positions significantly even with nearly unchanged activity.

To filter this noise from the actual signal, the author introduced a Residency stability mode. It calculates how often a node remained physically resident in VRAM after opening the tab or resetting the window. In a test with 18 Python code queries, 207 nodes were consistently resident, while in a poetry test, 193 nodes were. Both sets shared 169 common nodes.

The author concluded that the core of experts depends on the topic, but a large portion remains shared across different tasks. They also noted that the cache requires warming up within a single domain: alternating topics resulted in a smaller stable core. The changes have been submitted as an issue for the main project, and a ready-to-use release is available in a fork pending maintainers’ approval.