Inference
EN
Kilo Code

PinchBench

An OpenClaw agent benchmark from Kilo that measures successful task completion across standardized real-world agent workflows.

ModelsModels with a result.10
Top resultBest result on this benchmark.95.7%Pokee-Isaac 28B
Top-3 spreadPoints from first to third.5.7 pts
YearYear of release.2026

Results

10 results
#ModelResultThe score from the source.
01Pokee-Isaac 28B95.7%
02Nemotron 3.5 Bolt 74B-A9B BF1690.3%
03Sage-A1.590.0%
03Nemotron 3 Ultra90.0%
05Nemotron 3.5 Bolt 74B-A9B87.8%
06Nemotron 3.5 Lightning 30B A3B NVFP483.4%
07LFM2.5-24B-A2B83.2%
08LLaDA2.2-flash81.7%
09LFM2.5-2.6B68.2%
10LLaDA2.2-mini62.3%