# CSAT-benchmark-smol Small open-weight language models take the 2026 Korean college entrance exam, the 수능. Totals out of 450. | Model | Build | Weights (GB) | Korean (100) | Math (100) | English (100) | History (50) | Inquiry (100) | Total (450) | | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | Gemma 4 12B | QAT 4-bit | 7.16 | 95.5 | 98.7 | 97 | 33 | 75 | 399.2 | | Gemma 4 12B (w/o Thinking) | QAT 4-bit | 7.16 | 79 | 76 | 92 | 35 | 45.5 | 327.5 | | Ternary Bonsai 2 27B | | 8.6 | 86.5 | 98.7 | 94 | 27 | 79 | 385.2 | | Ternary Bonsai 2 9B | | 3.46 | 65.5 | 90.7 | 90 | 20 | 54.5 | 320.7 | | Gemma 4 E4B | QAT 4-bit | 6.14 | 57.5 | 81.3 | 90 | 38 | 50 | 316.8 | | Gemma 4 E4B (w/o Thinking) | QAT 4-bit | 6.14 | 58 | 68 | 90 | 25 | 42.5 | 283.5 | | Nanbeige4.2 3B | FP8 | 5.19 | 57 | 86.7 | 93 | 17 | 50.5 | 304.2 | | LFM3 1.7B | 8-bit | 1.84 | 45 | 86.7 | 76 | 18 | 35.5 | 261.2 | | Gemma 4 E2B | QAT 4-bit | 4.34 | 42 | 68.3 | 90 | 27 | 32 | 259.3 | | Gemma 4 E2B (w/o Thinking) | QAT 4-bit | 4.34 | 39 | 66.7 | 77 | 30 | 25 | 237.7 | | Qwen3.5 2B | 4-bit | 1.94 | 47.5 | 59 | 78 | 18 | 32 | 234.5 | | Kanana-2 1.3B | BF16 | 2.58 | 28.5 | 47 | 58 | 30 | 29.5 | 193 | | LFM3-VL 2B | 4-bit | 1.59 | 16.5 | 59 | 75 | 21 | 5.5 | 177 | Updated 2026-11-18. Memory, for scale: Galaxy Watch8 2 GB, iPhone 15 6 GB, iPhone 16 Pro 8 GB, RTX 5060 8 GB, Galaxy S26 12 GB, RTX 5070 12 GB, RTX 5080 16 GB, RTX 4090 24 GB, RTX 5090 32 GB. The defences proposed against AI in the exam hall work on the network: jamming, Wi-Fi shut off, radio detectors. A model of this size needs no network. Each one here fits in the memory of a laptop, and the smallest fit in a phone. Each model answers every question in its own request. The answers are graded and totalled out of 450 the way CSAT-benchmark does it. [Method](https://github.com/MercuriusDream/csatbench-smol) [CSAT-benchmark](https://github.com/hehee9/CSAT-benchmark) [@mercuriusdream](https://mercuriusdream.com/)