Nimbus 1.1 Prime-EE, measured and published
The first model of AhoosAI’s own. The scores, what moved against the same weights with the adapter off, and the one area that lost ground — with the weights on Hugging Face.
Nimbus 1.1 Prime-EE is a behaviour adapter over
Qwen2.5-Coder-7B-Instruct. It is not trying to make the
model a better programmer — the base is already a good one. It is
trying to fix what a good programmer gets wrong by habit: a hardcoded
dir, a form control with no label, a container that runs
as root.
The numbers
52.3 on IFEval, 61.6 on HumanEval, 67.2 on MBPP, all run with lm-evaluation-harness. Against the same weights with the adapter switched off, frontend went from 9/20 to 15/20 and backend from 14/20 to 15/20. Image specifications went the other way, from 18/20 to 15/20 — the one area with no training data behind it, which is where the cost of deferring that data shows up.
Every figure, how it was taken and what it is comparable to is on the benchmark page.
The weights
The adapter is published with its card, its training record and its regression stated plainly: AhoosAI/nimbus-1-1-prime-ee.