TLDR: Quantized versions of Llama 3.2 1B and 3B models with "competitive accuracy" to the original versions (meaning some degraded performance; plots included in the release notes).
They were already pretty small but I guess the smaller the better as long as accuracy doesn't suffer too much.