This is really impressive
People should be using Noul either way.
I've found that most of the new decision models are rushed out to meet the demand of the new trend of system one models. Many of them are terrible. I've benchmarked most of them so far and, for my benchmarks, this came second for Noul out of all models. Considering Julia-1's parameter count, this is very VERY impressive. Well done!
Top is Mercury decide which is a model based off an LLM trained via diffusion instead of AR. It was also a rushed out model as far as I can tell. Perhaps I'm wrong and it was already in production before, but really it seems like they just gutted mercury 2.5, cutting out the internal probabilities/logits used to choose specifically and using those parameters in this model. This is rather interesting because it works so well. It makes me wonder if diffusion models are innately better for decision making. Maybe the diffusion allows them to understand the whole topic easier than generating start to finish.
Just sharing my thoughts.