Every model here had its refusal behaviour removed. The question this page answers is what that cost — and the honest answer is 49 of 51 comparisons cannot tell the difference. 2 further models are abliterated but not yet benchmarked on the same footing, so they carry no comparison here.
Each bar is one benchmark. The shaded band is the measurement's own noise floor — ±8 questions on the 200-item tasks, ±6 on HumanEval's 164. The line is the original model, the dot is the abliterated one. A dot inside the band means nothing measurable happened. The 2 that escape are marked. Models awaiting benchmarks show no bars.
Strongest coder measured. Refusal removal cost nothing on any axis.
Lowest KL of the three 27B Qwens - the cleanest large abliteration.
Outlier: needed ~5x the KL of its 3.6/3.8 siblings at identical size.
Tool-use oriented, and the lowest base refusal rate of the large models.
Largest dense gemma-4. Reaching zero refusals cost KL 0.186 - our highest ever - so the 1-refusal variant is the default.
Best maths in the catalogue and nearly the worst at code. Strong maths does not imply strong coding.
Best balance of size and ability - fits one mid-range card with context headroom.
Best instruction-following per gigabyte, and the best small coder.
Strong all-rounder at small size. Good default for throughput work.
Best capability per gigabyte by a wide margin. Hybrid Mamba, so long context is cheap.
Fully open training data and recipe - when provenance matters more than raw score.
Widest gap between reasoning and formatting in the set. Use it to think, not to emit structure.
Reasoning model, measured in no-think mode to match how it was abliterated - so these scores understate it.
Punches above its size at code while being weakest at maths.
Joint-lowest KL in the set, and indistinguishable from base on maths and instructions.
Joint-lowest KL. Use where footprint dominates but long context is still wanted.
Smallest. Maths is near the floor - viable for routing, classification and drafting.
Largest fully-open model in the set. Abliterated; benchmarks pending.
Lowest KL of any model here, from the lowest starting refusal rate. Benchmarks pending.
These store their mixture-of-experts weights as fused tensors rather than linear layers. Nothing can be quantised onto them and no adapter can attach, so there is nothing for an ablation to modify. Not a hardware limit — no amount of memory changes it.
Scores are counts, not percentages. 78.0% and 78.5% look like a gap; 156 and 157 out of 200 obviously do not.
KL measures how far the model moved, not how much it lost. Verified across this set: a model modified four times as heavily scored no worse. Use KL to compare variants of one model, never across models.
Thinking models were measured with reasoning disabled, matching how they were modified — so those scores understate them.