CompareToolsMethod @niklaslenz_ai
Local AI · measured at home

I run it at home
so you don't
have to guess.

Which local AI should you run on your hardware? I test setups on my own machines with the same tests. Every number links to the raw output and the exact command.

Find your setup ↓
32complete runs
6memory classes with a run
2,488raw files public
6memory classes

Not up to date yetThe newest models are not on the site yet. I'm testing them now.

My test bench at home: a rack with a Mac Studio and four GB10 boxes between two GPU towers.
Photo: my setup. Background made with AI.
Your hardware

How much memory do you have for AI?

Pick your memory first. A class shows a run only when every part of it is done and checked.

Last measured Nov 1, 2026

16 GB 1× RTX 5060 Ti1 complete runAlsoimage 52.9 tok/s →
32 GB 1× RTX 5090 or 2× RTX 5060 Ti9 complete runsAlsoimage 253.4 tok/s →
128 GB 1× GB10 box9 complete runsAlsoimagevideo 64.6 tok/s →
256 GB 2× GB10 boxes5 complete runs 70.6 tok/s →
256 GBMac 1× Mac Studio4 complete runsAlsoimage 152.5 tok/s →
512 GB 4× GB10 boxes4 complete runs 81.8 tok/s →
Latest tests

New results

Each tile is one agent test: the model built a 3D scene on its own, with one attempt and no help. Click a tile for the recipe, the exact command and every number.

The page this setup built in the agent test512 GBDeepSeek-V4.1-Flash-1017NVIDIA NVFP4 checkpoint, TensorFold TP4 with DSpark drafts (Mia recipe @fcef6cf, NVFP4=1)87.7 overall (as of 1 Nov 2026) · 9 / 10 agent test · 69.7 tok/sThe page this setup built in the agent test256 GBnewQwen3.8-Flash-NextMia's dual recipe (NVFP4, TensorFold Zig, 2 GB10 boxes)71.4 overall (as of 30 Oct 2026) · 9.5 / 10 agent test · 58.3 tok/sThe page this setup built in the agent test128 GBnewKAT-Coder V2.5-DevCommunity NVFP4 build with grafted MTP head, vLLM 0.24.0 (howtospark recipe), 4 slots, 262k context65.0 overall (as of 26 Oct 2026) · 8.5 / 10 agent test · 64.6 tok/sThe page this setup built in the agent test512 GBnewDeepSeek-V4.1-Flash-1017Official checkpoint FP8/MXFP4, TensorFold TP4 with DSpark drafts (Mia recipe @6beaad9, 8 switches on)90.7 overall (as of 1 Nov 2026) · 9 / 10 agent test · 69.9 tok/sThe page this setup built in the agent test32 GBQwen3.8-27BEXL3 4.00 bpw, ExLlamaV3 1.6.0 (Mia kit, 2 GPUs), MTP49.0 overall (as of 25 Oct 2026) · 8 / 10 agent test · 70.8 tok/sThe page this setup built in the agent test128 GBQwen3.8-Flash-NextEXL3 3.05 bpw + MTP, ExLlamaV3 + TabbyAPI, MiniCPM 4.7 beside it (ViC305 recipe)55.4 overall (as of 26 Oct 2026) · 8.5 / 10 agent test · 56.2 tok/s
How I test

Five tests. No vibes.

Speed v2
tok/s

How fast the answer appears: tokens per second with one request, and the time to the first token at 2K, 8K and 32K.

Quality v2
7 tests

Fixed question lists with a margin of error next to each number: MMLU-Pro, GPQA, IFEval, LiveCodeBench, long text and GPQA with thinking.

Memory test
context

The largest context that loads and answers, the memory it needs and the decode speed there.

Build test
0 to 10

The model builds one 3D scene as an agent, with one attempt and no help. A browser checks the page.

sparkDash
Mia's tool

Mia's own benchmark tool, run unchanged, in its own column. Never mixed with my numbers.

Checked, not assumed

  • A run is on this site only when every part is done: Speed v2, all seven Quality v2 tests, the memory test, the build test and sparkDash. Every number has its raw file, and an independent check passed.
  • The exact software version is pinned for every run.
  • If an answer gets cut off by the length limit, it counts as wrong. The setup page says how many were cut off.
  • A run is ranked only after I have rated the page it built. The overall score (speed 35 %, half with one request and half with 8 requests at once, quality 25 %, browser check 5 %, my rating 35 %) decides the rank in each class. In a class with two or more ranked runs, the top run gets the label “Rank 1”. How it is made.