Minebench

MineBench is a free, web-based AI evaluation platform that tests spatial reasoning and instruction-following by having models generate 3D voxel builds from natural language prompts. Instead of relying on text or image outputs, models must output raw JSON block coordinates, which are then rendered in a Minecraft-inspired voxel world. Human voters compare two anonymous builds side by side, choosing the better one, and results feed a live Elo leaderboard. The platform offers two modes: Arena for head-to-head model comparisons and Sandbox for experimenting with prompts to see how different models perform. MineBench is designed for AI researchers, developers, and enthusiasts who need a transparent and reproducible way to assess model capabilities in spatial reasoning. Each build uses a grid of 1,247 blocks, and voters can rotate, pan, zoom, and orbit to inspect builds from any angle. The benchmark includes a curated set of prompts that probe various spatial concepts, from simple structures to complex scenes. The voting interface also provides 'Skip' and 'Both bad' options, ensuring fair comparisons. Key features include real-time leaderboard updates, textures from the Faithful Pack (Minecraft-inspired), and no post-processing or image generation—pure block output. The platform is entirely web-based, free, and inspired by MC-Bench and VoxelBench. MineBench is ideal for comparing models like GPT, Claude, and Gemini on a specific capability that traditional text-based benchmarks may not capture. Unlike general-purpose benchmarks, MineBench focuses solely on spatial reasoning via voxel construction, making it a niche but valuable tool for evaluating 3D understanding. Its lack of API or batch processing keeps it accessible but not scalable for enterprise use. For researchers and hobbyists, it provides a fun and intuitive way to probe model strengths and weaknesses.