Local LLM quote-drafting evals (Mac mini, Ollama)
Collection
Tests, graded answers and speed behind Staffbox.ai: an AI on a Mac mini that drafts quotes from your price list. Fictional test shops. • 6 items • Updated
date stringdate 2026-09-23 00:00:00 2026-10-10 00:00:00 | machine stringclasses 4
values | model stringclasses 10
values | streams int64 1 4 | per_stream_tok_s float64 4.6 172 ⌀ | wall_for_220_tokens stringlengths 3 11 ⌀ | notes stringclasses 10
values |
|---|---|---|---|---|---|---|
2026-09-23 | Mac mini M4, 16 GB | qwen3:8b via Ollama, default parallelism | 1 | 20.6 | 65 s | includes model load |
2026-09-23 | Mac mini M4, 16 GB | qwen3:8b | 2 | 20.4 | 112 s | requests queued |
2026-09-23 | Mac mini M4, 16 GB | qwen3:8b | 4 | 20.5 | 127 s | requests queued |
2026-10-01 | Mac mini M4, 16 GB, freshly erased, Staffbox installer | qwen3:8b via Ollama 0.35.0, flash attention, KV cache q8_0 | 1 | 20.2 | 11.2 s | model already warm |
2026-10-01 | Mac mini M4, 16 GB, freshly erased, Staffbox installer | qwen3:8b | 2 | 20.3 | 21.8 s | requests queued (one at a time) |
2026-10-01 | Mac mini M4, 16 GB, freshly erased, Staffbox installer | qwen3:8b | 4 | 20.3 | 43.5 s | requests queued |
2026-09-23 | Mac mini M4, 16 GB | qwen3:8b at 64K context, Hermes Agent one-shot task | 1 | null | 5 to 10 min | read a vault file, answer, write a log line; browser also running |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3:8b via Ollama, default parallelism | 1 | 62.2 | 37 s | includes model load |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3:8b | 2 | 114.4 | 4 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3:8b | 4 | 113.8 | 8 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | gpt-oss:20b, 64K context | 1 | 111.5 | 74 s | includes model load |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | gpt-oss:20b, 64K context | 2 | 138.8 | 3 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | gpt-oss:20b, 64K context | 4 | 139.8 | 7 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3:30b-a3b, 64K context | 1 | 154.6 | 79 s | includes model load |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3:30b-a3b, 64K context | 2 | 172.2 | 3 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3:30b-a3b, 64K context | 4 | 171.5 | 5 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3.8:27b, 64K context | 1 | 30.5 | 83 s | includes model load |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3.8:27b, 64K context | 2 | 46 | 10 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3.8:27b, 64K context | 4 | 44.7 | 21 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | gpt-oss:120b, 8K context | 1 | 14.5 | 19 s | model already loaded; split 36% GPU / 64% CPU; cold load from HDD took 12 min |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | gpt-oss:120b, 8K context | 2 | 14.5 | 31 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | gpt-oss:120b, 8K context | 4 | 14.5 | 62 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3-coder-next, 8K context | 1 | 35.9 | 8 s | model already loaded; split 45% GPU / 55% CPU |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3-coder-next, 8K context | 2 | 35.6 | 13 s | requests queued |
2026-10-06 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | qwen3-coder-next, 8K context | 4 | 35.6 | 26 s | requests queued |
2026-10-07 | Dell Precision 5820, Xeon W-2133, 80 GB DDR4, RTX 3090 24 GB | MiniMax-M2.7 UD-Q2_K_XL via llama-server | 1 | 4.6 | null | not scripts/bench.py: llama-server chat endpoint, 400-token answer; 22.7 GB on GPU, rest in RAM |
2026-10-10 | Dell Precision 5820, Xeon W-2133, 128 GB DDR4 (8 x 16 GB, all 4 channels), RTX 3090 24 GB | qwen3-coder-next, 8K context | 1 | 43.4 | 7 s | model already loaded; split 45% GPU / 55% CPU |
2026-10-10 | Dell Precision 5820, Xeon W-2133, 128 GB DDR4 (8 x 16 GB, all 4 channels), RTX 3090 24 GB | qwen3-coder-next, 8K context | 2 | 43.1 | 11 s | requests queued |
2026-10-10 | Dell Precision 5820, Xeon W-2133, 128 GB DDR4 (8 x 16 GB, all 4 channels), RTX 3090 24 GB | qwen3-coder-next, 8K context | 4 | 43 | 22 s | requests queued |
2026-10-10 | Dell Precision 5820, Xeon W-2133, 128 GB DDR4 (8 x 16 GB, all 4 channels), RTX 3090 24 GB | gpt-oss:120b, 8K context | 1 | 16 | 15 s | model already loaded; split 36% GPU / 64% CPU |
2026-10-10 | Dell Precision 5820, Xeon W-2133, 128 GB DDR4 (8 x 16 GB, all 4 channels), RTX 3090 24 GB | gpt-oss:120b, 8K context | 2 | 16 | 28 s | requests queued |
2026-10-10 | Dell Precision 5820, Xeon W-2133, 128 GB DDR4 (8 x 16 GB, all 4 channels), RTX 3090 24 GB | gpt-oss:120b, 8K context | 4 | 16 | 56 s | requests queued |
2026-10-10 | Dell Precision 5820, Xeon W-2133, 128 GB DDR4 (8 x 16 GB, all 4 channels), RTX 3090 24 GB | MiniMax-M2.7 UD-Q2_K_XL via llama-server | 1 | 7.3 | null | not scripts/bench.py: llama-server chat endpoint, 400-token answer; same settings as the 7 Oct 80 GB row |
Staffbox bench
Ollama inference speed for a local LLM (qwen3:8b) on a 16 GB Mac mini M4, in dated rows.
By Staffbox.ai, an AI worker on a Mac mini that drafts quotes from your own price list. Measurements on Staffbox's own Mac mini only, no customer data; no customer has run Staffbox yet. [email protected]
Dated measurements of local LLM inference speed on the hardware Staffbox runs on, taken with scripts/bench.py. Rows are added, never edited.
Status, 10 Oct 2026: speed measured on two machines, a Mac mini M4 16 GB and a Dell Precision 5820 with an RTX 3090 24 GB (at 80 GB and 128 GB of RAM). More RAM speeds up only models too big for the GPU: qwen3-coder-next 35.9 to 43.4 tok/s, gpt-oss:120b 14.5 to 16.0, MiniMax-M2.7 4.6 to 7.3. Ollama serves one request at a time, so concurrent requests queue. Rows are in measurements.csv; per-answer times for the quote desk (average 1.1 to 2.9 s on the 2 Oct runs) are in the scorecards.
Accuracy lives in staffbox/staffbox-scorecards.
To contribute a row, run bench.py, then open an issue on Staffbox-ai/staffbox with the machine, model and output. Built from docs/measurements.md at commit 86343d9.
- Downloads last month
- 88