LocalLLMAdvisor/local-llm-advisor-benchmarks

Local LLM Advisor: measured llama.cpp benchmarks

624 llama.cpp measurements (llama-bench, plus a few llama-server chat-timing rows from the Strata comparisons) taken by Local LLM Advisor on rented NVIDIA GPUs, covering 12 GPUs (rtx-3060-12 / rtx-3060, rtx-3090, rtx-4060, rtx-4060ti-16, rtx-4070, rtx-5060ti-16, rtx-4090, rtx-5070, rtx-5070ti, rtx-5080, rtx-5090, rtx-pro-6000), 14 GGUF models and context depths of 0, 4096, 16384, 32768, 65536, 131072 tokens (depths above 16384 only for the RTX 3090, RTX 4090 and RTX 5090, Q4_K_M, models that fit, plus Qwen3.8 27B Q6_K at 32768 on the RTX 3090 and RTX 4090, and Q8_0 at 32768 on the RTX 5090). Every row is our own run; no third-party benchmark rows are included in this dataset.

own_measurements_multigpu.csv (config multigpu) holds 21 further rows from a dual RTX 3090 box (device_count 2, hw_id 2x-rtx-3090, llama.cpp default layer split across both GPUs, CUDA 12.8, b11658, depths 0/4096/16384): six Q8_0 models of 25-37 GB plus Llama 3.3 70B Q4_K_M (about 19 tok/s generation). They are kept separate from the single-GPU file and are not used for the site's per-card medians.

Attribution (CC BY 4.0)

This dataset is licensed under CC BY 4.0. When you use it, attribute it as:

Data: Local LLM Advisor (localllmadvisor.com)

The full leaderboard (which also aggregates public community results under their upstream licenses) is at https://localllmadvisor.com/leaderboard.html, with a CSV download at https://localllmadvisor.com/data/leaderboard.csv. Source and tooling: https://github.com/mattgrandy/local-llm-advisor.

Method

• Tool: llama-bench from llama.cpp build b11658 (commit 3e4dfa8f3), CUDA 12.8 build, flash attention on, all layers on GPU (-ngl 99), 3 repetitions (-r 3), single GPU.
• pp_tps = prompt processing tokens/s (pp_n prompt tokens, default 512); tg_tps = generation tokens/s (tg_n tokens, default 128). Rows with depth > 0 measure speed with that many tokens already in the context.
• Rented machines differ in CPU and RAM; per-row details (driver, CPU, GGUF file, standard deviations) are in the notes column.
• Dates of runs: 2026-10-16, 2026-10-17, 2026-10-18, 2026-10-21.

strata_measurements.csv (config strata) holds 43 rows measured with the Strata engine (engine = strata, strata 0.2.7 / build 0.1.53) on 2026-10-28: an RTX 3090 (241.8 GiB host RAM), an RTX 3060 12GB and an RTX 5060 Ti 16GB (two runs; the first on a busy shared host, flagged in notes, the rerun on a less loaded host with its load average recorded in notes). Models: Qwen3.6 35B-A3B UD-Q4_K_XL and Qwen3.8 Flash Next (Q2_0 on the 3090, IQ3_S on the 3060 and 5060 Ti), MTP on and off, at empty context, realistic ~100-token chat prompts with 768-token replies, and 4K and ~35K prompts. Speeds come from the server's own timings, not from llama-bench, so they are not directly comparable with the llama.cpp file; each row records the launch command, engine flags, host RAM in use and MTP acceptance. The same Qwen3.6 GGUF was also run with llama.cpp b11658 on each card for comparison (rows in the default config, quant UD-Q4_K_XL, with any host-load notes).

Columns

column	meaning
hw_id	Local LLM Advisor hardware id
device_raw	GPU name as reported by the driver
device_count	number of GPUs used
backend	llama.cpp backend (CUDA)
model, quant	model name and GGUF quantization
params_b, size_gib	parameters (billions) and GGUF file size (GiB)
fa, ngl, threads, ubatch	flash attention, GPU layers, CPU threads, micro-batch
depth	tokens already in context before the measurement
pp_n, pp_tps	prompt tokens and prompt-processing speed
tg_n, tg_tps	generated tokens and generation speed
build, build_num	llama.cpp commit and build number
n_samples	repetitions averaged by llama-bench
date, notes	run date and free-text run details
engine	inference engine (llama.cpp for every row in the default config)

Caveats

Single runs of one machine per GPU model on rented hosts: treat numbers as indicative. Speeds depend on driver, CPU, thermal and power limits.