Kumulipo build-off

One prompt, two machines. The same request for a scroll-driven motion graphics piece about the Kumulipo went to a 27B open model on a garage GPU server and to Claude Opus 5.5. Both built it inside Claude Code, with no human help.

Time and resources

Builds in progress. Check back in a few minutes.

See the output

Pick a build to load it here.

Scroll inside the frame. The chant text is left as placeholders on purpose: models invent Hawaiian lines that look real, so the real text will come from Liliʻuokalani's 1897 translation, which is public domain.

The prompt

Loading…

The garage server

Garage GPU server

Self-hosted by Purple Maia on dedicated NVIDIA GPUs

Model: Qwen3.8 27B, dense (open weights)
Quantization: NVFP4 and 6-bit builds, load-balanced, with multi-token prediction
Context window: 262,144 tokens
Concurrency: About 8 requests at once
Speed, 1 request: 83 tokens/sec
Speed, 8 at once: 588 tokens/sec total (74–83 each)
API: OpenAI- and Anthropic-compatible gateway (LiteLLM)
How Claude Code reached it: ANTHROPIC_BASE_URL pointed at the gateway

Speeds measured 13 Nov 2026 with 400-token responses.