Advisor / Blog
Blog
Posts built from our own measurements and open dataset. Every number is computed from the data files.
State of local LLM hardware, October 2026
2026-10-17 · 253 llama-bench rows on 6 GPUs plus 858 public rows across 24 hardware entries: what fits per memory tier, how much speed long context costs, MoE vs dense and Q8 vs Q4.
Data is CC BY 4.0, see using our data.