MoE A3B Fine-tune · 35B Total / 3B Active

Qwopus3.6-35B-A3B-v1

Same 17-prompt suite as the Qwopus3.6-27B v1-preview eval — 5 agentic + 1 nothink rerun, 5 web-design, 6 canvas/WebGL. Q5_K_M on a single RTX 5090 via llama.cpp. Headline: 162 tok/s avg, 129k tokens generated, 13.3 min runtime. The MoE is 2.6× faster than the dense 27B preview at a larger quant. The web-design outputs are some of the best one-shot HTML I've seen out of any open model in this size class — verbose in a good way, the pages feel complete on the first attempt where most models land surface-level.

i
A real upgrade for one-shot front-end on a 5090 Per the model card, this is a 9%-trainable LoRA fine-tune over Qwen3.6-35B-A3B with three-stage curriculum SFT. The MoE speedup (3 B active of 35 B) is a massive practical win, and an updated Qwopus3.6-27B is in the works that should carry similar improvements forward.
161.9avg tok/s×2.6 vs 27B preview
14shipped runs
106,688completion tokens
~25 GBVRAM used
65Kcontext window

Web design · open to preview

SaaS landing pagePrism — AI observability
75.9 KB · 23,839 tok · 151 s
Analytics dashboardLight theme, emerald accent
37.5 KB · 14,029 tok · 87 s
Designer portfolioMaya Chen — kinetic typography
27.5 KB · 9,144 tok · 56 s
Pricing page3 tiers + animated toggle + FAQ
50.1 KB · 13,862 tok · 86 s
Mobile app marketingStillwater — CSS-only iPhone mock
47.9 KB · 16,596 tok · 104 s

Canvas / WebGL · creative coding

3 of 6 creative-canvas runs shipped clean and are below. The Mandelbulb shader, soft-body physics sandbox, and audio-reactive visualizer didn't render correctly on first attempt — these are prompts most one-shot models in this size class fail on, and they're typically the kind of brief that needs a second turn to fix shader errors or collision math. Excluded here for the headline; raw outputs are still in the repo if you want to inspect.

Particle attractor3000-particle fluid swarm
10.6 KB · 4,152 tok · 25 s
Generative flowfieldInk-line agents on simplex noise
19.3 KB · 6,934 tok · 42 s
Three.js crystal sceneTransmissive glass + bloom
16.1 KB · 5,674 tok · 35 s

Agentic reasoning · text output

Multi-step planningURL shortener deploy plan
thinking: 2,440 tok · 15 s
Self-critique loopPalindrome · O(n³) → O(n²)
thinking: 4,391 tok · 27 s
Code debug (4 bugs)k-th smallest element
thinking: 1,393 tok · 9 s
Structured JSON extractionCalendar + roster from prose
thinking: 2,501 tok · 15 s · clean pass
JSON extraction · no-thinkSame prompt, thinking off
352 tok · 2 s
Tool-use planningWeather + flights + hotel
thinking: 1,381 tok · 8 s