Language model memory
Local LLMs need memory for model weights, context and runtime buffers. Quantisation helps, but it does not make memory irrelevant.
Images are different
Image generation can place a very different load on the GPU, especially with identity conditioning or higher resolutions.
Voice and 3D still count
STT/TTS and a live 3D avatar may be lighter individually, but simultaneous workloads can matter more than any single benchmark.
Why AVRE will benchmark combinations
A useful compatibility page should test the real product stack rather than assuming a GPU model tells the whole story.
This article describes AVRE’s current product direction and general technical reasoning. Details can change as development and testing continue.