bare/gpt-oss-20b / tier-1-cube-with-hole

success tier 1

bare (openai/gpt-oss-20b)

render

loading STL…

success 1.5s · 143/106t

SCAD (5 lines)

difference() {
    cube(50, center = true);
    translate([0, 0, -25]) // extend below the cube to ensure full cut
        cylinder(d = 20, h = 100, $fn=64, center = false);
}

```openscad
difference() {
    cube(50, center = true);
    translate([0, 0, -25]) // extend below the cube to ensure full cut
        cylinder(d = 20, h = 100, $fn=64, center = false);
}
```

meta

runId	`bare_gpt-oss-20b-e01a1a89770a-2026-04-29T07-28-54-312Z`
matrixId	`bare/gpt-oss-20b`
signature	`e01a1a89770aee2b…`
provider	openai-self-hosted
model	openai/gpt-oss-20b
harness	bare
status	success
duration	1.51s (wallclock: load + prompt eval + generation + network)
TTFT	0.138s
generation time	0.68s (pure decode, excludes load + prompt eval)
tokens	143 in / 106 out
generation tok/s	156.3 tok/s (pure decode rate)
wallclock tok/s	70.2 tok/s (end-to-end)
createdAt	2026-04-29T07:28:54.313Z
gitCommit	`68749000bf3a`
openscad	OpenSCAD version 2026.04.27

model metadata (self-hosted)

セルフホスト LLM はクラウドの単一 provider と違い、実質「(モデル開発者, weight publisher, 量子化)」の組で挙動が決まる。 run 時に LM Studio / Ollama から取得した値:

publisher	`openai`
type	llm
arch	gpt-oss
quantization	`MXFP4`
max ctx	131,072 tok
capabilities	tool_use
host GPU	NVIDIA GeForce RTX 3090 (Vulkan)
host VRAM	24 GB
host CPU	12th Gen Intel(R) Core(TM) i9-12900K
host RAM	32 GB

prompt

OpenSCAD で、50mm 角の立方体の中央に、直径 20mm の貫通穴を z 軸方向に開けたモデルを作成してください。

完成したコード全体を ```openscad ... ``` のフェンスで囲んで出力してください。コードのみで、追加の説明は不要です。