MineBench says one prompt generated 100M-block 3D NYC at 8,192³ grid scale
BorisMPower · x · 2026-09-08
Benchmark outfit MineBench claims its Astra Pro model handled an extreme test: with gridSize raised from 256³ to 8,192³ (32,768x more voxel positions), a single prompt asking for an accurate, to-scale New York City produced a 61 KB program generating 100.1 million blocks — 1.65 GB as JSON, spanning roughly 8.2 km at 1 m per block, too large for their renderer to fully display. The reposter muses that the whole world may be rendered in 3D this year. Note: the "GPT-6 Astra Pro" name is unverified.
More from Multimodal
- fal brings native real-time 1080p video generation to H3 Max: 5s video in under 5s — jfischoff · 2026-09-08
- Newcomer shocked: ComfyUI tutorials are AI-generated and the Stable Diffusion Discord is dead — RawkShalk · 2026-09-08
- Phone-Shoot Mixed with 3D and V2V: An AI Video Workflow Worth Stealing — mrjonfinger · 2026-09-08
- TMNT Fan Short Made with MiniMax H3 Shows Off Its Video Generation — darthfurbyyoutube · 2026-09-08
- Fan-Made AI Trailer Imagines Game of Thrones Season 9 — Suibeam · 2026-09-08
- "Initialize": An AI-Generated Music Video — VIV-AF-D · 2026-09-08