shinhipc/local-ai-benchmarks

HighLife PC — Local AI Field Benchmarks (October 2026)

🇰🇷 한국어 설명: README_ko.md

HighLife PC is a computer shop in Cheonan, South Korea. We run several local AI servers for day-to-day store work, and this dataset collects numbers we measured ourselves, on our own machines, with our own settings — not vendor spec sheets. A few vendor/third-party figures are included only for context and are clearly marked in the source_type column as 외부(...) ("external").

• Rows: 273 (249 own measurements, 24 external reference values). Ids run M001–M279; the 6 ids removed from the public version are left empty.
• Period: 2026-10-12 to 2026-10-27
• Topics: LLM inference speed (prefill, decode, concurrency), power & temperature, quantization, long context & conversation cache, accuracy (GSM8K, in-house items), coding-agent scores, image generation speed, Korean embedding retrieval
• Text fields are mostly Korean; metric_key is an English identifier. A short glossary is below.

Hardware

Code	Configuration	Main workload
DEV-A	RTX PRO 5000 48GB ×1 · Core i7-14700KF · DDR5 128GB · PCIe 5.0 x16 · Ubuntu	Qwen3.8-Flash-Next (MoE) on Strata
DEV-B	RTX 5090 32GB ×2 · Threadripper 2950X (X399) · DDR4 128GB · PCIe 3.0 x16 · Linux	27B chat model (NInfer) + image generation; dual-GPU Strata trial
DEV-C	ASUS Ascent GX10 (NVIDIA GB10) ×4, switchless ring	DeepSeek-V4.1-Flash on SGLang (TP=4)
DEV-D	RTX 5090 32GB ×1 (also drives the display) · 128GB RAM · Windows	Quantization comparison on 10-16; 10-12 agent tests (placement estimated). Configuration no longer exists
DEV-E	RTX PRO 5000 48GB in DEV-B chassis, PCIe 3.0 x8 slot	Configuration at the time of the 10-12 agent tests (estimated)

DEV-B1 = one RTX 5090 of DEV-B, DEV-Bcpu = CPU of DEV-B only, EXT = not our hardware (external values or cloud models).

Key results

1) One 48GB card + 128GB RAM serving Qwen3.8-Flash-Next (DEV-A, Strata 0.1.50, final production config on 2026-10-27)

IQ4_XS, 524K context, 2 parallel slots, 300W power limit. Speed and the 479K read were measured with a 6GB conversation cache; the resume test was run later the same day after raising the cache to 20GB and the engine memory limit to 120GB.

Metric	Value
Decode, 1 user	150–179 tok/s
Decode, 2 users (sum)	204 tok/s
Prefill 11K / 22K	4,900–4,980 / 5,260–5,380 tok/s
Read a 479,096-token document	104.5 s (4,599 tok/s), hidden note found
Resume a 480K-token conversation (cache hit)	0.8–1.6 s (first read 105–109 s)

2) Aggregate decode vs. number of concurrent users (short prompts)

Machine · engine · model	1	2	4	8
DEV-A · Strata · Flash-Next IQ3_S (1M × 4 slots)	150	171	230	(out of RAM)
DEV-B 5090×2 · Strata · Flash-Next IQ3_S (parallel 8)	179	185	243	317
DEV-B1 5090×1 · NInfer (new) · 27B with MTP	127 (Korean)	252	–	608
DEV-C GX10×4 · SGLang · DeepSeek-V4.1-Flash	65.6	–	145.8	–

Models differ in size and architecture — read this as "how each setup scales with users", not as a hardware ranking.

3) GPU power limit (DEV-A, 2026-10-23, mixed full load for 15+ min per step)

Limit	Short-prompt decode	Prefill sum under load	Avg power	Max temp
250W	143.7	3,923	239.9W	82°C
285W	140.3	4,113	265.8W	85°C
300W	142.9	3,863	275.0W	87°C

No difference under mixed load, but for a single long prefill (16K) 300W was ~6% faster than 250W (4,900–5,085 vs 4,600–4,710 tok/s, 10-27).

4) GSM8K by quantization (DEV-A, 200 problems, thinking on)

Quant	Correct	Wall time
IQ3_S (official)	197/200 (98.5%)	476 s
IQ4_XS (Orca derivative)	196/200 (98.0%)	539 s
UD-IQ4_XS (Unsloth)	195/200 (97.5%)	572 s

5) Coding agents — same model, different harness (self-made test, 35 problems, scaled to 100, 2026-10-23)

Model	Harness	Hard 5	General 30	Avg 35
DeepSeek-V4.1-Flash (DEV-C)	Hermes Agent	84.1	97.7	95.8
DeepSeek-V4.1-Flash (DEV-C)	DSH 0.2.1-rc.2	93.9	98.7	98.0
DeepSeek-V4.1-Flash (DEV-C)	Pi 1.3.0	96.3	98.5	98.2
Qwen3.8-Flash-Next IQ3_S (DEV-A)	DSH	84.7 (2nd run 95.9)	99.4	97.3
Qwen3.8-27B (DEV-B1)	DSH, PYTHONUTF8 off → on	36.8 → 90.4	–	–
Claude Opus (cloud, reference)	Claude Code	96.3	98.9	98.5

This is a self-made test; the problems are not released — only scores. Re-running the same setup moved scores by about ±5 points, so 1–2 point gaps should be treated as ties.

6) Image generation (DEV-B1 RTX 5090, ComfyUI 0.42, 1024×1024)

Model	Sec / image	Peak VRAM	Korean text (2 lines)
FLUX.3 image klein (NVFP4)	2.6–2.9	~10GB	2/2 correct
Edit: klein vs Qwen-Image-Edit-2511	5.6–5.8 vs 19.1	10.9 vs 29.6GB	–

Columns

Column	Meaning
id	Row id (M001…)
group	Test group name (Korean)
date	Measurement date; external rows show e.g. (공식 발표) = "official announcement"
device / device_spec	Machine code and public specs
engine	Inference engine and version/commit
model / quant	Model and quantization
settings	Key settings (context, parallel slots, power, cache, …)
metric_key / metric_ko	Metric id (English) / Korean label
value	Value exactly as recorded (a~b range, 197/200 fraction, 2.49→0.44 before→after)
value_min / value_max	Numeric columns, filled only for single numbers or ranges
unit	Unit (초 = seconds, 점 = points, 문항/문제 = items)
method	One-line method (Korean)
source_type	자체측정 own, 자체측정(클라우드 모델) own run on a cloud model, 외부(공식 발표) vendor, 외부(공개 보고) public report, 외부(타인 측정) third-party doc, 외부(공개 자료) public data
notes	Caveats, corrections, context (Korean)

Methodology

• LLM speed: our own Python scripts streaming to OpenAI-compatible endpoints, thinking off, temperature 0.6.
  • prefill tok/s = input tokens ÷ time to first token (includes request overhead)
  • decode tok/s = output tokens ÷ time after first token
  • N concurrent = N identical short prompts in parallel, per-user speeds summed
  • Some rows use engine-log values (see method). With speculative decoding (MTP etc.) client-side chunk counts under-count tokens, so engine logs were preferred there.
• Power test: per step 3 short prompts, 3 × 2,000-token answers, 15+ min of 2-user full load, 3 min rest between steps, fixed case fans.
• GSM8K: first 200 problems of the public test split (openai/grade-school-math), thinking on, temperature 0.6, max 6,144 tokens, 2 parallel. The number on the "Answer:" line (else the last number) is compared; failed requests count as wrong.
• In-house quality items: 69 items (chat 20, tool calls 15, knowledge/reasoning 14, lawful-but-sensitive 17, long docs 3); rule-based grading + blind grading by another local model.
• Coding agents: self-made test, 35 tasks (not released), same instructions, automated tests + human review, one attempt per task. The 25- and 30-task agent tests of 10-12 are also self-made.
• Images: fixed seed, identical English scene prompts, per-image time excluding the first (model load), peak per-process VRAM from nvidia-smi.
• Embedding retrieval: 51 store product/info docs, 40 customer-style Korean queries, top-1 / top-3 hit rate.

Limitations

• Do not mix vendor claims with our measurements. Rows whose source_type starts with 외부 are context only; we did not reproduce them.
• Small samples: mostly 1–5 runs; ±5–10% run-to-run variation was common (the 4-node ring's code decode ranged 57–111 tok/s, so medians of 5 were used).
• Rows often differ in model, quant, engine version and context — always read settings.
• These are production servers; some rows may include background traffic (noted in notes).
• Coding/agent/quality tests are self-made and not released; only scores are.
• Some models are community derivatives (e.g., refusal-direction removed); results may differ from the base model.
• Where our logs contain a correction, only the corrected value is included; for repeated identical measurements the latest value is used.
• The 10-12 agent tests ran on an earlier hardware layout; part of that placement is estimated from our logs (marked "추정" = estimated in notes).
• Some image-model rows were removed from the public version.

License

CC-BY-4.0 — commercial reuse allowed with attribution. Please cite as "HighLife PC Local AI Field Benchmarks (2026-10)". Model and engine licenses follow their respective distributors.

Contact

HighLife PC (Cheonan, South Korea) · Instagram @hipc.ai