ISA-Bench (First Attempt): leaderboard

Metric: Tasks solved (%; share of the 92 tasks of five programming games with constrained instruction sets (SIC-1, Human Resource Machine, TIS-100, MHRD, BOX-256) whose program passes every verification case in the game's own parser, virtual machine and verifier; first attempt only, no worked solutions; Ollama default builds, temperature 0.7, 4,096 output tokens). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 14 models tracked.

Top models

#ModelScore
1ISA-Bench Gemma-4 31B (Thinking, Ollama build)25
2ISA-Bench Qwen3 14B (Thinking, Ollama build)15.22
3ISA-Bench Qwen3 32B (Thinking, Ollama build)14.13
4ISA-Bench Qwen3 8B (Thinking, Ollama build)14.13
5ISA-Bench DeepSeek-R1 32B (Ollama build)13.04
6ISA-Bench Devstral-Small-2 (Ollama build)13.04
7ISA-Bench Qwen3-Coder 30B (Ollama build)13.04
8ISA-Bench Qwen2.5-Coder 32B (Ollama build)10.87
9ISA-Bench Granite-4.1 8B (Ollama build)6.52
10ISA-Bench Qwen2.5-Coder 14B (Ollama build)3.26

Interactive version: theaggregate.ai/benchmark?slug=isa-bench-first-attempt · How It Works · Data refreshed daily, snapshot 2026-09-26.