Model briefing

Mixture of experts: active parameters are only part of the story

Gemma 4 26B-A4B is a useful example of why compute-oriented model labels and a local memory budget answer different questions.

What the source record tells us

Sparse text and image model with about 3.8B active parameters.

The reviewed catalog entry records roughly 25.8 billion stored parameters and 3.8 billion active parameters. Activating a subset for a token does not make the remaining model weights disappear from an all-on-device setup. Offloading changes the execution plan and needs its own performance check.

What this means in practice

When comparing sparse models, record both stored and active parameters. Use stored weights for the planner’s all-on-device memory estimate, then test speed and answer quality separately on your intended workload.

What this briefing does not establish

This briefing interprets the model-card and configuration records already reviewed for the 99Tokens catalog. It does not report a new release date, independently measured benchmark or fresh availability check. Runtime support and the exact file you use still need checking.

FOLLOW A CONNECTION

Where this can take you

Explore all topics