LLM benchmarks · v3 testpack · deterministic · no API key

Benchmark LLMs on YOUR hardware.

The same 34 deterministic tasks, scored entirely on your machine — no judge model, no API key, one 0–100 score. A catalog of 109 models, ranked by runs people actually made on their own hardware. The only public LLM board that ranks where the model runs, not just which model it is.

Every score here is a real run. We seed nothing — which is why the board is small and why you can trust the numbers on it.

$ npx @pipelinescore/cli

That's the whole command. It finds your Ollama / LM Studio / llama.cpp / MLX server, lists your models, auto-detects your hardware, and walks you through the rest.

Run on your hardware Rank the rigs →See where you'd rank →

Models tracked 109Real runs 24Users 5Testpack v3Tasks 34Scored locally · server never re-scores

The model board

Best run per model. Click a column to re-rank, pick two rows to go head-to-head.

How scores are computed →

Weighting

13 models · sorted by Balanced composite (high → low) · pick any two rows to compare

#	Model	PipelineScore ▾	Code	Reason	Tool Use	RAG	Speed	Runs	Tier
1	openai/gpt-oss-20blocal	93.2	100.0	80.0	87.5	100.0	99.3	2	TRUNK
2	gpt-oss:latestlocal	91.6	100.0	80.0	87.5	100.0	88.1	1	TRUNK
3	qwen3.6-35b-a3blocal	87.5	87.5	80.0	87.5	100.0	84.2	4	MAINLINE
4	qwen3-coder:30blocal	86.6	100.0	70.0	100.0	100.0	54.5	1	MAINLINE
5	qwen3.8:27blocal	86.2	100.0	90.0	87.5	100.0	37.9	1	MAINLINE
6	qwen3.8-27b@iq3_xxslocal	84.2	87.5	80.0	87.5	100.0	62.1	5	MAINLINE
7	mlx-community/Qwen3.6-35B-A3B-4bitlocal	83.7	75.0	80.0	87.5	100.0	82.5	1	MAINLINE
8	/home/thomaskyn/models/Qwen3.6-35B-A3B/Qwen3.6-35B-A3B-UD-IQ3_S.gguflocal	80.9	100.0	80.0	87.5	100.0	17.3	1	MAINLINE
9	gemma4:12b-it-qat_gpulocal	80.9	87.5	80.0	87.5	100.0	40.6	1	MAINLINE
10	qwen/qwen3.6-35b-a3blocal	79.3	75.0	60.0	87.5	100.0	82.6	2	MAINLINE
11	qwen2.5-coder:7blocal	77.7	100.0	70.0	87.5	100.0	10.2	2	MAINLINE
12	/home/thomaskyn/models/Qwen3.5-4b/Qwen3.5-4B-Q4_K_M.gguflocal	57.8	50.0	40.0	87.5	87.5	29.3	2	TAP
13	/home/thomaskyn/models/Qwen3.5-9b/Qwen3.5-9B-Q4_K_M.gguflocal	50.5	25.0	60.0	87.5	75.0	12.0	1	TAP

Popular matchups

The rivalries worth settling. Every pair opens a live head-to-head.

openai/gpt-oss-20b vs gpt-oss:latest gpt-oss:latest vs qwen3.6-35b-a3b qwen3.6-35b-a3b vs qwen3-coder:30b qwen3-coder:30b vs qwen3.8:27b qwen3.8:27b vs qwen3.8-27b@iq3_xxs

m5-max-48gb vs ryzen-7-7700-cpu-48gb ryzen-7-7700-cpu-48gb vs rtx-5080-16gb rtx-5080-16gb vs rtx-5070-ti-16gb rtx-5070-ti-16gb vs m3-ultra-96gb

Five measures. One number.

Code is executed, reasoning is exact-match, tool use and RAG are JSON-match, speed is measured throughput. No judge model, no rubric, no API key.

Code

28%

Reason

22%

Tool Use

18%

RAG

17%

Speed

15%

The tiers

Every score maps to a tier, named the way pipelines are: from TRUNK (top of the network) down to DRIP.

TRUNK

90–100

MAINLINE

75–89

FEEDER

60–74

TAP

40–59

DRIP

0–39

Hardware board

Which rig wins?

Every hardware tag ranked by its best run — Apple Silicon vs consumer GPUs vs datacenter cards vs CPU-only.

Users board

Every run. Every user.

The community board — sortable, filterable by provider, tier, and hardware. Find someone who ran your model on your rig.

STEP 01

Point the CLI at your model

Ollama, LM Studio, MLX, llama.cpp — anything OpenAI-compatible. Local runs need no account and no API key.

STEP 02

Tag your hardware

--hardware-tag m3-max-128gb / rtx-4090-24gb / a100-80gb. Same model on different rigs gets ranked separately — that's the point.

STEP 03

Land on the board

A deterministic 0–100 PipelineScore across 34 tasks, computed on your machine, plus a tier badge and a public spot on the hardware-aware leaderboard.