holistechlabsneun Modelle
DEEN

Measured · 15 June 2025 · three runs per model

Nine models on the HP ZBook Ultra G1a

dense against mixture of experts

Authors

Dipl.-Ing. Sören GebbertInstitut für holistische Technologieforschung GmbH Natalia GebbertInstitut für holistische Technologieforschung GmbH

Machine

HP ZBook Ultra 14 inch G1a

SoC
AMD Ryzen AI Max+ PRO 395
GPU
Radeon 8060S
Memory
128 GB – dedicated graphics memory set to 512 MB in the BIOS
OS
Windows 11 Pro 24H2

Runtime – the same for every model

LM Studio 0.3.16

Engine
llama.cpp Vulkan v1.34.1
Offload
as many layers on the GPU as the machine allowed
Context
8,192 tokens, identical for every run
KV-Cache
on the GPU

Task

Summarise

Source
English Wikipedia article on Plato, 27 KB
Prompt
a two- to three-sentence introduction plus four bullet points
Runs
three per model: one without cache, two with
Models
9 – from 0.6 billion to over 100 billion parameters

Where this report comes from

Article from 2025

Published
15 June 2025
Evidence
results table and 47 screenshots from LM Studio
This edition
newly typeset, charts computed from the same figures
Successor
the measurement of 8 September 2026 on the same machine

What this machine did in 2025

44.0Token/s

Qwen3-0.6b Q8 the fastest model in the field

25.8Token/s

Qwen-30B-A3B Q4_K_M a factor of 5.6 over the equally large QwQ 32B Q8

0.24seconds

wait with cache instead of 119.4 s without, on QwQ 32B Q8

4.3Prozent

widest spread across three runs; every other model less

Speed figures are means over the three runs of each model. The wait without cache comes from the first run, the one with cache is the mean of the two that follow. Every figure appears in the results table of the original article.

01 · The field

Nine models, one machine

Between the fastest and the slowest model lies a factor of 9.5: 44.0 against 4.6 tokens per second. Both ran on the same machine, on the same task.

Nine models, one machine Output speed in tokens per second, mean of three runs per model. HP ZBook Ultra 14-inch G1a · Ryzen AI Max+ PRO 395 · Radeon 8060S · 128 GB LM Studio 0.3.16, llama.cpp Vulkan v1.34.1 · 8,192-token context window, KV cache on the GPU · measured 15 June 2025 dense experts Qwen3-0.6b Q8 44.0 t/s Qwen-30B-A3B Q4_K_M 25.8 t/s Qwen-30B-A3B Q8 23.4 t/s Qwen3-4B Q8 19.8 t/s Nous-Hermes-2-Mixtral-8x7 Q8 10.7 t/s Gemma-3 12b 7.7 t/s Magistral-Small Q8 6.6 t/s Llama-4-Scout-17B-16e-Instruct 6.1 t/s QwQ 32B Q8 4.6 t/s Both 30-billion expert models beat every dense model above 4 billion.
One exception stands out. Llama-4-Scout-17B-16e-Instruct is a mixture of experts as well and still sits in the bottom third. Chapter 05 says why.

02 · Build

Why the experts come out ahead

A dense model computes every token using all of its parameters. A mixture-of-experts model stores just as many but computes with only a part. Memory demand follows the first number, speed follows the second.

Two models in this field make the difference immediate, because their names state the same order of magnitude. Qwen-30B-A3B Q4_K_M is a mixture of experts and reaches 25.8 tokens per second. QwQ 32B Q8 is dense, about the same size – and manages 4.6. That is a factor of 5.6 in favour of the build, at practically the same parameter count.

The relationship is not without exception. Llama-4-Scout-17B-16e-Instruct is a mixture of experts as well and stays behind several dense models at 6.1 tokens per second – for a reason that has nothing to do with the build. Chapter 05 shows it.

A year later this became the basis of the next measurement: Qwen3.8-Flash-Next holds 512 experts and activates ten per token. What shows up here as an advantage carries a 176.9-billion-parameter model on the same notebook there.

03 · The cache

What is left the second time

The same prompt, asked a second time: 119.4 seconds of waiting become 0.24. A factor of 487 – and the output speed stays the same throughout.

What the cache does to the wait Seconds to the first token, logarithmic. Open circle: without cache, filled: with cache. HP ZBook Ultra 14-inch G1a · Ryzen AI Max+ PRO 395 · Radeon 8060S · 128 GB LM Studio 0.3.16, llama.cpp Vulkan v1.34.1 · 8,192-token context window, KV cache on the GPU · measured 15 June 2025 0.05 s 0.1 s 0.5 s 1 s 5 s 10 s 60 s 120 s QwQ 32B Q8 Llama-4-Scout-17B-16e-Instruct Magistral-Small Q8 Qwen-30B-A3B Q4_K_M Qwen-30B-A3B Q8 Nous-Hermes-2-Mixtral-8x7 Q8 Gemma-3 12b Qwen3-4B Q8 Qwen3-0.6b Q8 Up to two minutes of waiting become fractions of a second.
The wait belongs to the prompt, not to the model. Once the processed prompt sits in the cache it all but disappears – equally so in every size class.

For actual use this is the difference between workable and unworkable. The first call on a long text makes you wait; no follow-up question on the same text does. Working through a document, you pay for the processing once rather than with every question.

04 · Reproducibility

Asked three times, the same number three times

No model spreads beyond 4.3 percent across its three runs. Seven of nine stay under two.

Three runs, the same number Output speed of the three runs per model, as deviation from that model's mean. HP ZBook Ultra 14-inch G1a · Ryzen AI Max+ PRO 395 · Radeon 8060S · 128 GB LM Studio 0.3.16, llama.cpp Vulkan v1.34.1 · 8,192-token context window, KV cache on the GPU · measured 15 June 2025 dense experts -4 -2 0 2 4 Llama-4-Scout-17B-16e-Instruct 4.3 % Qwen-30B-A3B Q8 2.4 % Gemma-3 12b 2.0 % Qwen3-0.6b Q8 1.7 % Qwen-30B-A3B Q4_K_M 1.6 % Qwen3-4B Q8 1.6 % Nous-Hermes-2-Mixtral-8x7 Q8 1.0 % Magistral-Small Q8 0.9 % QwQ 32B Q8 0.6 % Spread No model spreads beyond 4.3 percent; seven of nine stay under 2.
The figures hold. The widest spread belongs to Llama-4-Scout-17B-16e-Instruct, the one model that did not fit entirely onto the GPU.

05 · The limit

Where the machine stopped short in 2025

Eight of the nine models ran entirely on the GPU. The ninth could not: in 2025 the Vulkan driver released at most 64 GB of graphics memory, and Llama-4-Scout needed more. Six of its 48 layers stayed on the CPU. The sensors show what that costs: the GPU clock falls from 2,650 to 846 MHz and utilisation to 72 percent while the CPU joins in at 61 percent.

The 2025 article closed with the assumption that a driver update might lift this limit. The measurement of 8 September 2026 on the same machine answers that, if under a different operating system: on Linux the kernel parameter amdgpu.gttsize releases 110 GiB of system memory to the GPU, and a 176.9-billion-parameter model runs. So the limit sat in the software stack, not in the machine.

06 · Every figure

The complete results table

27 runs, exactly as they appear in the original article. Every chart in this report is computed from this table.

Model Run Tokens/s Tokens Wait (s) Cache
Qwen3-0.6b Q8144.474852.26no
243.714460.06yes
343.784290.06yes
Qwen-30B-A3B Q4_K_M125.6666867.12no
225.809470.08yes
326.077730.09yes
Qwen-30B-A3B Q8123.1266662.42no
223.686510.08yes
323.518360.08yes
Qwen3-4B Q8119.6162710.09no
219.856970.09yes
319.925980.09yes
Nous-Hermes-2-Mixtral-8x7 Q8110.6520838.46no
210.752250.13yes
310.762660.13yes
Gemma-3 12b17.7532428.61no
27.652890.17yes
37.603330.17yes
Magistral-Small Q816.6092480.20no
26.562,5610.19yes
36.621,0410.22yes
Llama-4-Scout-17B-16e-Instruct15.9230182.54no
26.113560.23yes
36.182860.21yes
QwQ 32B Q814.65782119.42no
24.641,1250.24yes
34.621,2970.25yes

07 · Limits

What this work does not show

No quality. What was measured is speed, not the quality of the summaries. Whether one model solves the task better than another is not stated here.

One prompt, one length. Every run uses the same text with an 8,192-token context. The measurement says nothing about behaviour on very long contexts – that is what the 2026 report is for.

One stack, one moment. Windows 11, LM Studio 0.3.16, llama.cpp Vulkan v1.34.1, as of June 2025. Drivers and engines have changed since; the figures hold for this stack and this date.

Three runs. Enough to show the spread, too few for a statistical claim about rare outliers.