smolmaxxing adapts local models to older, low-memory devices

Move model data between storage and memory

The project streams selected model components from storage and uses custom computation kernels. Its device-specific vocabulary reductions require modified model files and runtime code; ordinary llama.cpp builds cannot load every supplied variant.

smolmaxxing device experiments ↗

Small memory use still needs storage and setup

The iPad experiment requires a modified device. Phone model files can exceed twelve gigabytes. The published speeds describe the maker’s particular hardware; they do not establish interactive quality or speed on another device.

smolmaxxing device experiments ↗

What you can do

1. Read the primary description and its stated limits before planning an evaluation. Open reference ↗

Where the evidence stops

Source text checked; no device modified or model executed. Independent discussions do not verify the advertised device measurements.

Sources read. Public source text checked November 14, 2026 UTC. No third-party project, model inference or full workflow executed.

Sources & further reading