Model Evaluation

Ornstein-Hermes-3.6-27B — SABER · Q5_K_M

eval by Kyle Hessling · model by @DJLougen at GestaltLabs

Agentic reasoning, production-grade front-end design, and canvas / WebGL creative coding. 14 runs, 71,587 completion tokens, self-hosted on a single RTX 5090. Refusal-shaped (SABER) Hermes 3.6 finetune over Qwen 3.5 27B.

Read the full report →
56.7avg tok/s
14runs
71,587completion tokens
~24 GBVRAM used
16Kcontext (f16 KV)

Web design · open to preview

SaaS landing pagePrism — AI observability
34.5 KB · 9762 tok · 172 s
 Analytics dashboardLight theme, emerald accent
23.4 KB · 7707 tok · 144 s
 Designer portfolioMaya Chen — kinetic typography
9.3 KB · 5862 tok · 103 s
 Pricing page3 tiers + animated toggle + FAQ
20.3 KB · 5534 tok · 97 s
 Mobile app marketingStillwater — CSS-only iPhone mock
14.1 KB · 8728 tok · 154 s

Canvas / WebGL · creative coding

Particle attractorCursor-attracted particle swarm
5.5 KB · 2282 tok · 40 s
 Generative flow fieldPerlin-driven agent motion
10.9 KB · 5649 tok · 99 s
 Three.js crystal sceneTransmissive glass + bloom
11.8 KB · 6839 tok · 120 s
 Audio-reactive visualizerMic + oscillator fallback
10.7 KB · 6352 tok · 112 s

Agentic reasoning · text output

Code debug (4 bugs)k-th smallest element
no-think: 1039 tok · 18 s
 Multi-step planningURL shortener deploy plan
no-think: 8000 tok · 140 s
 Self-critique loopPalindrome · O(n³) → O(n²)
no-think: 1521 tok · 27 s
 Structured JSON extractionCalendar + roster from prose
no-think: 1200 tok · 21 s
 Tool-use planningWeather + flights + hotel
no-think: 1112 tok · 20 s