Minebench

MineBench is a free, browser-based platform for evaluating how well AI models understand and follow spatial instructions. Instead of relying on multiple-choice tests or human raters scoring text, MineBench asks models to convert natural-language prompts into raw JSON block coordinates. Those coordinates are rendered as 3D voxel constructions in a Minecraft-inspired scene, giving you a direct visual read on a model's capacity for spatial reasoning and instruction-following. Two complementary modes keep the platform useful for different tasks. Arena mode throws two models into a head-to-head match: you see both builds side by side and vote on which better satisfies the original prompt. Every vote updates a live Elo leaderboard, so you can watch rankings shift in real time. Sandbox mode lets you type any prompt you want and generate a build from a single model, which is ideal for quick experimentation or stress-testing a specific capability. The builds themselves are confined to a grid of 1,247 blocks using Faithful Pack textures, and you can orbit, zoom, and pan to inspect every angle. To keep voting fair, MineBench includes options to skip a match or mark both builds as bad, and it explicitly positions itself as a pure block-output benchmark—no image generation, no post-processing. It draws inspiration from MC-Bench and VoxelBench but adds a human-voting layer that many automated benchmarks lack. MineBench is best suited for AI researchers, developers, and enthusiasts who want a hands-on, transparent way to compare models like GPT, Claude, and Gemini on a narrow but meaningful dimension: 3D spatial understanding. It's not a general-purpose benchmark, and it currently offers no API, batch processing, or paid tiers. But for interactive evaluation and community-driven ranking, it's a lightweight, accessible option that fills a real niche.