4x3090 Benchmarks: 27B Q8KXL at 32GB Beats Next IQ3 at 82GB on Both Speed and Security Tasks
Repulsive_Initial308 · reddit · 2026-08-28
A user running llama.cpp on 4x RTX 3090s shared initial tests of the 3.8 Next IQ3 quantization, calling it "meh." Their existing mix of Q4/Q8/BF16 already feels near-perfect. Early results show 3.8 27B Q8KXL at 32GB of weights is vastly superior to 3.8 Next IQ3 at 82GB — worse in security vulnerability assessment AND slower. He notes it's early days, so things may change.
More from Infra
- Open-source Discord AI assistant Zauq: Multi-model routing & Docker sandbox — rar_file-exe · 2026-08-28
- Utility proposes rate cuts due to data center profits, saving users $100/year — bennash · 2026-08-28
- Why your local LLM feels dumber than hosted versions: Implementation details — JeremyCMorgan · 2026-08-28
- GTA VI Size Rumored to Range Between 0.748TB and 2TB — PtrPomorski · 2026-08-28
- Own a frontier AI model running locally in just 5 hours — MaziyarPanahi · 2026-08-28
- Optimal parameter settings for local deployment of Qwen3.8 Flash Next shared — Motor_Ad16 · 2026-08-28