Nimbus 2 Apex: one gain, and everything that did not move
Instruction-following up 8.7 points against the model it was trained on, and no loss anywhere that survives a significance test. The thinking ladder did not work, and the reason is in the data.
Nimbus 2 Apex is a LoRA adapter on Qwen3-8B: 87 million trained parameters, 1.05% of the model, on 36,024 rows and 106 million tokens. It was measured against the weights it was trained on, item by item, in the same session and on the same loaded model — so the difference between the two arms is the adapter and nothing else.
It follows instructions better. IFEval goes from 70.7 to 82.0, with 26 questions fixed against 9 broken, which a coin would do about six times in a thousand. Nothing else moved in a way that survives the same test: MMLU-Pro, AIME and LiveCodeBench all shifted by less than the churn underneath them.
Two things are worth reading the page for. The adapter first scored 10.7 on MMLU-Pro —
chance, on a ten-choice benchmark — because it ends its answers with
\boxed{\text{B}} and the pattern that read boxed answers refused any brace
inside them. It had forgotten nothing. And the ladder, the thing the model was built
around, did not work at all: it writes 1,983 words at level 20 where the base model writes
1,979. The level was measured from the reasoning it labelled, so the model could predict
the text exactly as well by ignoring it.
What ships is the adapter at three quarters strength. At full strength it writes longer than the base model, the ceiling cuts it off mid-thought, and it loses ten points at level 5. Turned down, the verbosity goes and the loss with it, and the instruction-following gain stays.