8 hours, same web-dev prompt: 9 models tested on an RTX 3060 12GB
zyxciss · reddit · 2026-09-20
- The author ran the exact same web-dev prompt (a polished NOVA//LABS single-page site, explicitly avoiding typical AI-slop UI) across 9 models on an RTX 3060 12GB / 16GB RAM Arch Linux setup via llama.cpp, recording all generations.
- Frontier references: GPT-5.6 Sol (1m06s, fastest), Gemini 3.8 Flash via Antigravity (3m12s, 9K tokens), Claude Sonnet 5 (4m56s).
- Local standouts: Bonsai 2 27B Ternary (7.66GB, 34–36 tok/s, 45 min, 52K tokens used out of 122K context) and Qwen 3.8 27B GSQ-RCO-IQ3-XXS + MTP (10.4GB, 29–40 tok/s, context compacted twice, 57 min), the latter called the most interesting local model.
More from Models
- Reddit user: GPT 6 reasoning feels worse than 5.6 Terra, like GPT 3 flashbacks — FigInternational4873 · 2026-09-21
- Google launches Gemini 3.8 Flash and Flash Cyber: 54.9% HLE, 2.6x better vuln patches — AlienX__0 · 2026-09-21
- Open-source model Jev called out for not crediting Lava, its 8192-context prior work — nischay_twt · 2026-09-21
- Devin makes SWE-2 free across Cloud, CLI and Desktop until October 8 — HankYeomans · 2026-09-21
- Encoder-style classification gets hot again: one multimodal BERT solved 1,000+ classification tasks — cwolferesearch · 2026-09-21
- Deployers told: benchmarks are a directional signal, not a substitute for your use case — evijit · 2026-09-21