[clbrmbr]

But it’s so much weaker than 4.5 in broader tasks… maybe more optimized against benchmarks but it’s just no replacement for a huge model.