Early-Preview Fine-tune · Round 2

Qwopus3.6-27B — v1-preview

by Kyle Hessling · built on Jackrong's fine-tune

Same 16-prompt suite as the Qwen3.6-27B base eval — 5 agentic, 5 web-design, 6 canvas/WebGL. Q4_K_M on a single RTX 5090 via llama.cpp. Byline metrics: 62.3 tok/s avg, 87.4 k tokens generated, 23.4 min runtime.

Read the full report PDF version Compare · Qwen3.6 base eval Follow @KyleHessling1

62.3avg tok/s+12.7% vs base

16runs

87,394completion tokens

~20 GBVRAM used

65Kcontext window

Web design · open to preview

SaaS landing pagePrism — AI observability

36.7 KB · 9,961 tok · 160 s

Analytics dashboardLight theme, emerald accent

37.4 KB · 13,190 tok · 213 s

Designer portfolioMaya Chen — kinetic typography

23.1 KB · 7,356 tok · 118 s

Pricing page3 tiers + animated toggle + FAQ

24.3 KB · 8,061 tok · 129 s

Mobile app marketingStillwater — CSS-only iPhone mock

29.3 KB · 8,005 tok · 128 s

Canvas / WebGL · creative coding

Particle attractor3000-particle fluid swarm

11.1 KB · 4,249 tok · 68 s

WebGL MandelbulbRaymarched fractal shader

11.5 KB · 4,364 tok · 70 s

Three.js crystal sceneTransmissive glass + bloom

17.9 KB · 6,375 tok · 102 s

Physics sandboxSoft-body collisions, fling mouse

15.1 KB · 4,384 tok · 70 s

Audio-reactive visualizerMic + oscillator fallback

12.0 KB · 3,018 tok · 48 s

Agentic reasoning · text output

Multi-step planningURL shortener deploy plan

thinking: 3,158 tok · 50 s

Self-critique loopPalindrome · O(n³) → O(n²)

thinking: 1,277 tok · 21 s

Code debug (4 bugs)k-th smallest element

thinking: 1,628 tok · 26 s

Structured JSON extractionCalendar + roster from prose

no-think rerun · 353 tok

Tool-use planningWeather + flights + hotel

thinking: 1,174 tok · 19 s

Model: Jackrong/Qwopus3.6-27B-v1-preview-GGUF — Q4_K_M · served via llama.cpp on a single RTX 5090