Qwen3.8:27B local port of 39k-line C file loses badly to Opus 5's 21-minute run
codehamr · reddit · 2026-08-24
The author tested the new Qwen3.8:27B on a brutal job: one-shot porting a 2.1 MB (600k token) single-file C game to single-file HTML/three.js. Setup: vLLM FP8, FP8 KV cache, full 262k context on an RTX 6000 Pro 96GB.
| Agent | Model | Wall clock | Lines out | Result |
|---|---|---|---|---|
| Claude Code | Opus 5 (cloud) | 21 min | 1759 | okay |
| hermes | Qwen3.8:27B | 4h18m | 949 | bad |
| codehamr | Qwen3.8:27B | 1h40m | 1056 | bad |
Only the Opus port reached playable quality. Key takeaway: local models live or die on the prompt — same weights under two very different harnesses produced the same broken port; a verbose harness doesn't rescue a thin prompt, it just burns GPU time. The wall-clock gap (hours of local GPU vs 21 cloud minutes) stands out most. Both the C original and the codehamr harness are open-sourced.
More from coding & agent
- "Vibecoding" is as misleading a label as "AI" and "SI," dev complains — eschadiol · 2026-10-16
- Doc2Harness: New benchmark finds best coding agent passes only 53% across 329 cases — _akhaliq · 2026-10-16
- Perplexity open-sources SPACE, the Firecracker microVM sandbox behind Perplexity Computer — perplexity_ai · 2026-10-16
- Blackstone goes all-in on Factory for autonomous software development — matanSF · 2026-10-16
- Longtime Windsurf user cancels $10/month legacy plan, citing AI subscription fatigue — tristanbob · 2026-10-16
- 'Dropping a 3,000-line PR twenty minutes after ticket assignment' — AI coding meme goes viral — facontidavide · 2026-10-16