SLM Performance Evaluation 9 benchmarks: HellaSwag, ARC-Easy, ARC-Challenge, PIQA, MMLU (continuation), WinoGrande, OpenBookQA, SciQ, CommonsenseQA Boris-2-Think · 49.53 avg Boris-2 · 48.78 avg GPT-X3-150M · 47.75 avg SmolLM2-135M · 45.99 avg KLM-1.6 · 48.07 avg BananaMind-2.5-Pro · 47.44 avg LFM-2.5-350M-Base · 52.64 avg 50.13 50.27 46.16 43.22 45.28 K 43.60 45.16 HellaSwag 63.93 59.76 59.64 58.63 61.07 K 62.79 68.86 ARC-Easy 34.81 32.00 31.31 29.69 33.53 K 31.14 41.13 ARC-Challenge 72.31 72.85 71.33 68.44 70.18 K 68.12 67.63 PIQA 29.30 28.96 28.87 29.90 29.26 K 29.24 33.32 MMLU (continuation) 53.67 52.57 54.62 52.96 52.33 K 52.57 55.25 WinoGrande 34.40 36.60 34.40 33.40 32.80 K 34.00 36.20 OpenBookQA 87.10 85.10 80.30 78.20 89.30 K 85.10 84.90 SciQ 20.15 20.88 23.10 19.49 18.92 K 20.39 41.28 CommonsenseQA Boris-2-Think, Boris-2, BananaMind-2.5-Pro were evaluated with lm-eval on all nine tasks. For GPT-X3-150M, SmolLM2-135M, KLM-1.6, LFM-2.5-350M-Base,HellaSwag, ARC-Easy, ARC-Challenge and PIQA come from the Open SLM Leaderboard and the other five tasks from lm-eval. The avg next to each name is the unweighted mean of all nine.LFM-2.5-350M-Base sits right of the dotted line as a larger-model reference. Panels are scaled independently to each task's top score.