See what local models actually built
Fixed prompts, versioned harnesses, and the output kept exactly as it came off the GPU. Filter the work, compare the same demo, then open the real artifact in your browser.
- demos
- 8
- models
- 13
- results
- 37
- coverage
- 8%
Fresh off the GPU
The most recent thing each model produced. Click one to see it full size, with what it cost to build.
Where the work is
Coverage by model and harness. Pick a row to use it as a demo filter.
Harness coverage
Browse the demos 8
Start broad, or narrow to the exact model and harness you want to inspect.
qwen3.8-flash-next-strataminimal-v1
qwen3.8-flash-nextminimal-v1
muse-glimmer-30bminimal-v1
qwen3.8-27bminimal-v1
gemma-4-31bminimal-v1
qwen3.6-27b-q6minimal-v1
qwen3.6-27bminimal-v1
nemotron-3.5-lightning-30b-a3bminimal-v1
qwen3.6-35b-a3bminimal-v1
Breakout
The canonical brick-breaker, fully specified. Everyone builds the same game, so the difference is entirely in feel, polish and whether the collisions actually work.
9 models
1 harness
9 results
qwen3.8-flash-next-strataminimal-v1
qwen3.8-27bminimal-v1
muse-glimmer-30bminimal-v1
qwen3.6-35b-a3bminimal-v1
gemma-4-31bminimal-v1
qwen3.6-27b-q6minimal-v1
qwen3.8-flash-nextminimal-v1
Bounce — Level 1
A momentum-driven puzzle-platformer vertical slice, with the level design left to the model. One bounce height, horizontal momentum only, an hour on the clock and five minutes at the end to make sure it still runs.
7 models
1 harness
7 results
qwen3.8-flash-next-strataminimal-v1
qwen3.8-flash-nextminimal-v1
qwen3.8-27bminimal-v1
gemma-4-12bminimal-v1
gemma-4-31bminimal-v1
qwen3.6-27b-q6minimal-v1
qwen3.6-27bminimal-v1
Snake, with one idea of your own
A specified Snake plus exactly one original mechanic the model has to invent, name and make work. The open-ended half is the interesting half.
7 models
1 harness
7 results
qwen3.8-flash-next-strataminimal-v1
qwen3.8-27bminimal-v1
qwen3.8-flash-nextminimal-v1
muse-glimmer-30bminimal-v1
gemma-4-31bminimal-v1
qwen3.6-27bminimal-v1
Bounce — the full game
Four levels played as one run, with lives and score carried through, two new surfaces introduced a level at a time and a Game Complete screen at the end. The level-1 demo asks whether a model can make the physics feel fair; this one asks whether it can finish a game.
6 models
1 harness
6 results
qwen3.8-flash-next-strataminimal-v1
qwen3.8-27bminimal-v1
Analytics dashboard
KPI tiles, two hand-drawn charts, a sortable table and a theme toggle — from a fixed dataset, so every model is drawing the same numbers.
2 models
1 harness
2 results
qwen3.8-flash-next-strataminimal-v1
qwen3.8-27bminimal-v1
Fix the WszystkieKina city schedule
A dense eight-day cinema timetable whose nested grids, sticky regions and filters fail in ways that tempt a broad rewrite. Tests structural CSS diagnosis and restraint.
2 models
1 harness
2 results
qwen3.8-flash-next-strataminimal-v1
qwen3.8-27bminimal-v1
Fix the platformer
A tiny platformer that runs but cannot be completed. Tests whether a model can read someone else's game loop, find real collision bugs, and then make it feel good.
2 models
1 harness
2 results
qwen3.8-flash-next-strataminimal-v1
qwen3.8-27bminimal-v1
Product landing page
A one-page marketing site with a hero, features, pricing and a footer. Tests layout, type, colour and restraint more than logic.
2 models
1 harness
2 results
No demos match these filters.
Clear one filter, or return to the full index.