Badass model
People complaining about reasoning verbosity are overlooking the fact that a model this small can even sustain reasoning for that long.
It's a fine model. Just wish it had MTP to make the latency more bearable.
It's a fine model. Just wish it had MTP to make the latency more bearable.
@ManIDontlike working on it, just a few days.
I asked this model to solve a basic mechanics problem involving a cylinder rolling down an inclined plane. It spent 64k tokens reasoning and still failed to produce an answer.
It's impressive that a 4B model can sustain such a long reasoning process, but this also highlights the difference between generating long chains of thought and actually converging on a solution. More reasoning tokens don't necessarily mean better reasoning.