Bench 2 crowns Nemotron 3 Nano

This is a useful benchmark, but the fixed token cap is clearly part of the story: it measures model quality plus how efficiently each model thinks within budget. The result still says something real about size-class specialization, especially at 3-4B.

• –Nemotron 3 Nano is the standout because it stays within budget and hits 100% on finance, which is exactly where many small models usually wobble
• –Phi-4 Mini looks like the best balanced generalist, with strong finance and code plus a much less lopsided profile than the others
• –Granite4:3B and Nemotron 3 Nano split into coder vs reasoner personalities, which is a strong argument for task-specific model selection at this size
• –Qwen 3.5 4B’s low score looks more like a truncation failure than pure capability loss, so per-model budgets are the right next experiment
• –The benchmark is most interesting as a methodology warning: fixed budgets can distort comparisons between thinking and non-thinking models