Asterisk (hybrid ASPP-attention SmolLM2)
NoesisLab/Asterisk is a researchy twist on the lightweight SmolLM2-135M-Instruct family: it adds an ASPP-style “parallel propagation” branch and fuses it with attention using a learned gate. The idea is to capture some benefits of local, iterative state evolution (which feels closer to graph/message passing) while keeping the global context modeling and token mixing that attention is good at.
If you’re tracking experiments in “make small models reason a bit better,” this is a nice example because the model card includes a concrete architectural sketch (ASPP steps, gating, and where it plugs into the decoder layer) plus a quick set of benchmark numbers (HellaSwag / ARC / PIQA / WinoGrande). It’s unlikely to be a drop-in replacement for production assistants, but it is a useful reference if you’re building custom blocks or want to A/B “pure attention” vs “attention + extra operator” at small scale.
Quick stats from the listing feed: pipeline: text-generation · 1 like.
Source listing: https://huggingface.co/models?sort=modified