DIY dual-RTX-5060-Ti rig benchmarks 12 local 4B-35B models with real wall-clock times

Careful-Report6526 · reddit · 2026-10-28

• Tired of synthetic cloud leaderboards, the author built a dedicated Linux test server: two RTX 5060 Ti cards (32 GB VRAM total) running vLLM in Docker.
• 12 current local models from 4B to 35B were run through complete benchmark datasets with no subsampling: IFEval, IFBench, ARC AGI 1, SimpleQA, SimpleQA Verified, SciCode, and vLLM bench serve. Lineup includes Qwen 3.5/3.6/3.8 (4B/9B/27B/35B A3B), Gemma 4 (E4B/26B A4B/31B), and GLM 4.7 Flash in NVFP4/AWQ quants.
• Every model was tested twice per benchmark — thinking on and off — with official sampling parameters, while tracking wall-clock time and tokens/s to compute a 0–100 time-efficiency score.
• Key finding: some models score well in 20–30 minutes while others burn 10 hours reasoning in circles on simple questions; MoE models like Qwen 3.6 35B A3B and Gemma 4 26B A4B finished in a fraction of the time while matching or beating dense 27B/31B models.
• Full sortable results at llmbenchmarks.io/local-gpu-benchmarks, aimed at 16 GB/32 GB setups; no ads or tracking, new benchmarks (Chess Puzzles) already running.

Original post →

More from Infra

• SK Hynix Q3 smashes estimates: revenue up 327.9%, operating profit up 632.1% YoY — firstadopter · 2026-10-28
• Oracle to Deploy Google TPUs Outside Google Cloud for the First Time From 2027 — rohanpaul_ai · 2026-10-28
• SK Hynix and Micron both expect memory demand to outpace supply through 2030 — BenBajarin · 2026-10-28
• Epoch AI chart shows US share of global compute capacity climbing to ~72% — SamuelAlbanie · 2026-10-28
• Liquid Inference serves 10B tokens/day, 88.7% of inputs are cache reads — AccBalanced · 2026-10-28
• Poll: UK Net Support for a Nearby Data Centre Sits at -13%, Only North East Positive — SamuelAlbanie · 2026-10-28

Infra channel →