What the teacher built

DeepSeek-V4.1-Flash working real Tokyo footprints and generic Japanese places through the build environment on two B300s, 20 to 21 September 2026. Left of each image: the exterior from the south-east. Right: the south half cut away, so floors, stairs and furniture show.

250passing included episodes
210at 0.9 or higher
231distinct passing tasks
265included episodes

Snapshot scope: 250 passing included episodes across 231 distinct tasks. Tasks include real buildings, open sites and generic places; repeated attempts are included. Many failed runs were omitted by the original gallery exporter, so 250 / 265 is not an overall success rate. Scores are offline structural heuristics, not visual-fidelity ratings. Select an image to open its full-size asset.

Five large buildings, reasoning effort high

Real footprints of 50 to 95 blocks a side, 105 to 194 blocks tall, each built twice (batch t3).

Same footprint, three builders

The three real buildings Opus built earlier. Opus is the bar; Gemma 4 12B is the untrained student.

Included episodes, best score first

The score is the verifier's: under 0.3 a gate failed (no way in, rooms cut off, floors missing). Builds cut off by a time cap were handed in afterwards from their best checked version.

Result
Effort
Kind
Size