I tested 9 LLMs on the exact same web-dev prompt for ~8 hours — RTX 3060 12GB results (Rate the best!)

I’ve spent basically the last 8 hours testing different models on the exact same web-development prompt, and I finally finished. The whole point of this nine-hour test was that which local model matches the frontier-level intelligence at size and could fit easily in an RTX 3060-like consumer card. My setup: GPU: RTX 3060 12GB RAM: 16GB DDR4, single-channel OS: CachyOS (Arch Linux) Local models were run through my local llama.cpp setup. Same prompt for every model. I recorded the generations so you can actually judge the websites yourself rather than relying on my description. Prompt Build a polished, production-quality single-page website for a fictional high-end technology studio called NOVA//LABS. Goal: make it look genuinely designed by a strong human frontend developer, NOT like generic AI-generated SaaS UI. Requirements: * Use plain HTML/CSS/JavaScript or React + Tailwind if you strongly prefer it. * Everything must run locally with minimal setup. * Create the entire project/files yourself. * No backend, authentication, database, or unnecessary complexity. * Responsive desktop + mobile layout. * Strong typography, spacing, hierarchy, subtle motion, and excellent visual composition. * Dark, sophisticated visual language with restrained use of gradients/glows. * Avoid the typical AI-slop look: no excessive rounded cards, giant gradient blobs, random glassmorphism, meaningless statistics, or generic "Empowering the future" copy. * Make the copy specific and believable. * Include: A striking hero section with a concise headline. A subtle animated visual representing an abstract computational system. A small selected-work/projects section. A concise capabilities section. A strong closing CTA/footer. * Add tasteful interactions such as hover states, scroll reveals, and subtle cursor/mouse effects where they genuinely improve the design. * Prioritize visual quality over feature count. * Use freely available CDN assets only if genuinely necessary; otherwise create visuals with CSS/SVG. * Keep the implementation reasonably small and understandable. Most importantly: make strong design decisions yourself. Do not explain your design choices before building it. Start by creating the project and finish with the exact commands needed to run it. (SELF CONTAINED HTML WITH JS AND CSS) I wanted to see what the models actually build, not just how well they explain code. The models 1. Gemini 3.8 Flash ~3 min 12 sec Used Antigravity and consumed roughly 9K tokens. This was one of the frontier-model reference points for the test. 2. GPT-5.6 Sol ~1 min 6 sec Token usage wasn't available to me. Extremely fast compared with the local models, so this was another useful frontier reference. 3. Claude Sonnet 5 ~4 min 56 sec Token usage wasn't available. Also included as a frontier reference. (I was only able to use Sonnet 5 as my Claude-Code Max subscription had expired) Local models 4. Bonsai 2 27B Ternary ~45 minutes Native ternary / ~2-bit model Model size: ~7.66GB Average generation: ~34–36 tok/s Context: up to roughly 102K Used ~52K tokens out of a 122K context during this run Qwen 3.8 27B GSQ-RCO-IQ3-XXS + MTP ~57 minutes Model size: ~10.4GB High thinking enabled ~29 tok/s around full context Around 40 tok/s with a much smaller/near-empty context Context used reached roughly 75K Context was compacted twice Available context for this particular run was around 49K after the relevant setup/limits this was probably the most interesting local result for me. 6. Qwen 3.8 27B Q4_K_M Unsloth Dynamic 3 2+ hours ~16.4GB model Q4_K_M High thinking enabled Full-context generation dropped to roughly 4 tok/s Context reached roughly 96K Obviously requires significant CPU/RAM offloading on a 12GB GPU Flags used : --jinja --reasoning-preserve -fa on -fit off -ngl 99 --override-tensor "blk\.([0-9]|[1-3][0-9]|4[0-5])\.ffn_.*=CPU" -ctk q4_0 -ctv q4_0 --gpu-layers-draft all --spec-type draft-mtp --spec-draft-n-max 2 -lv 4 --no-mmproj -np 1 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 --load-mode none --no-warmup -b 256 -ub 128 -c 98304 7. Qwen 3.8 27B GSQ-RCO-IQ3-XXS + MTP — thinking OFF ~12 minutes Same general Qwen 3.8 GSQ-RCO model, but this time I disabled thinking. It used roughly 12K tokens and produced the site dramatically faster. This was a particularly useful comparison because it shows how much the reasoning mode itself can affect local generation time. 8. Ornith 1 9B Q4_K_M ~2.4 minutes Model size: ~5.4GB ~74 tok/s Native context: up to 262K This generation only used around 2.6K tokens This is the speed monster of the local group. 9. Ornith 1.5 35 A3B Q6 ~30 tok/s Model size: ~22.4GB ~30 tok/s Context available for this run: around 128K Obviously heavily dependent on offloading because of the model size Quick summary # Model Approx. time Local? Generation speed 1 Gemini 3.8 Flash ~3:12 ❌ — 2 GPT-5.6 Sol ~1:06 ❌ — 3 Claude Sonnet 5 ~4:56 ❌ — 4 Bonsai 2 27B Ternary ~45 min ✅ ~34–36 tok/s 5 Qwen 3.8 27B GSQ-RCO-IQ3-XXS + MTP ~57 min ✅ ~29–40 tok/s 6 Qwen 3.8 27B Q4_K_M Unsloth Dynamic 3 2+ hrs ✅ ~4 tok/s at full context 8 tok/s at empty 7 Qwen 3.8 27B GSQ-RCO-IQ3-XXS, thinking OFF ~12 min ✅ — 8 Ornith 1 9B Q4_K_M ~2.4 min ✅ ~74 tok/s 9 Ornith 1.5 35 A3B Q6 — ✅ ~30 tok/s My personal take For local models specifically, the one that impressed me the most was Qwen 3.8 27B GSQ-RCO-IQ3-XXS. It hit a pretty interesting balance between: actual design quality coding ability context handling generation speed fitting within a 12GB GPU setup The Qwen 3.8 27B Q4_K_M Unsloth Dynamic 3 was also interesting from a quality perspective, but the speed penalty once you're deep into the context is huge. Bonsai 2 27B Ternary was also surprisingly usable given that it's a ~7.66GB ternary model. so its Qwen 3.8 27B Q4_K_M > Qwen 3.8 27B GSQ-RCO-IQ3-XXS > Bonsai 2 27B Ternary I've attached the screen recording showing the outputs. Especially interested in other RTX 3060 / 12GB setups ;0 If possible Someone please post down GPT-6-ASTRA's results if they have a codex subscription.