Ocean UI Model Runs

Same blank Vite + Three.js starter, isolated opencode model runs, static deployment bundle.

Isolation Method

Removed the earlier hand-written Codex implementation before running the model bench.
Used the same blank Vite + Three.js starter for every model.
Ran each model in its own work/ directory with opencode run --dir <model>/work.
Prompt told models to work only inside the current directory and not inspect parent or sibling runs.
No model run received another model's output.

opencode run \
  --model opencode-go/<model> \
  --agent build \
  --format json \
  --dir <run>/work \
  --dangerously-skip-permissions \
  --title "ocean-ui <model>" \
  "$(cat <run>/work/PROMPT.md)"

Results

Model	Exit	Build Recheck	Desktop WebGL	Mobile Layout	Source	Screenshot
`deepseek-v4-flash`	0	pass	pass	pass	`deepseek-v4-flash/work/src/`	screenshot
`deepseek-v4-pro`	0	pass	pass	pass	`deepseek-v4-pro/work/src/`	screenshot
`glm-5.1`	0	pass	pass	pass	`glm-5-1/work/src/`	screenshot
`glm-5.2`	0	pass	pass	pass	`glm-5-2/work/src/`	screenshot
`kimi-k2.7-code`	0	pass	pass	pass	`kimi-k2-7-code/work/src/`	screenshot
`qwen3.7-plus`	0	pass	pass	pass	`qwen3-7-plus/work/src/`	screenshot
`mimo-v2.5`	0	pass	pass	pass	`mimo-v2-5/work/src/`	screenshot
`mimo-v2.5-pro`	0	pass	pass	pass	`mimo-v2-5-pro/work/src/`	screenshot
`minimax-m3`	130	not run	not run	not run	starter unchanged	none

Pixel Check

Desktop screenshots were checked for non-empty center-canvas pixels.

Model	Center Non-Dark Ratio	Color Buckets
`deepseek-v4-flash`	0.908	130
`deepseek-v4-pro`	0.992	141
`glm-5.1`	0.843	34
`glm-5.2`	1.000	160
`kimi-k2.7-code`	1.000	224
`qwen3.7-plus`	1.000	33
`mimo-v2.5`	1.000	204
`mimo-v2.5-pro`	1.000	39

Notes

Mobile check used a 390x844 viewport. All 8 completed outputs had a canvas, visible controls, and zero horizontal overflow.
minimax-m3 stalled before editing source. It was stopped after remaining unchanged.
Browser console warning/error count was 0 for all 8 completed outputs.
npm run build was re-run after model completion for all 8 completed outputs and passed.