Four gamesStrix HaloCurrent settings
Four games on CachyOS and Strix Halo
Slay the Spire 2, Rocket League, ARC Raiders and Overwatch: FPS, 1% lows, frame times, memory, captured settings and downloadable measurements.
The numbers behind the videos
Benchmarks.
What ran. What broke. What changed.
My gaming and local AI tests, with the setup and results open for you to explore.
11 published reportsResults from my own hardware
Featured test · Coding agents
Qwen 3.8 Flash-Next on Halogen: 72 coding runs
Three harnesses. Six tasks. Four attempts at each. The full run history, including the failures.
Runs resolved
66/ 72
Halogen W4B · xhigh reasoning
One configuration on a 128 GB desktop.
Gaming on Linux
FPS, frame pacing and memory from my own game captures, with the scene and settings limits beside the numbers.
Coding agents
The same tasks, different harnesses and model servers. Inspect the individual runs.
Halogen W4B72 runs128 GB
Qwen 3.8 Flash-Next on Halogen: 72 coding runs
Pass rates, generation speed and individual run evidence across three coding harnesses, six tasks and four repeats on a 128 GB desktop.
Two campaignsDifferent configurations
Qwen 3.8 Flash-Next: Halogen vs llama.cpp
Two campaigns side by side: task outcomes, speed as context grows, and the configuration differences that limit the comparison.
llama.cpp IQ4_XS18 runs
Qwen 3.8 Flash-Next on llama.cpp: 18 coding runs
The earlier IQ4_XS campaign at xhigh reasoning, with three coding harnesses, six tasks and the evidence from each run.
Model size & quality
What fits in memory, how quickly it runs, and what survives a smaller file.
QuantizationLow vs high reasoning
GLM-5.3-Flash: quantization on 128 GB
A ladder of model sizes at low and high reasoning. See which files fit, which tests finish, and the raw model outputs.
First runOne-shot tests
GLM-5.3-Flash: the smallest quant on 128 GB
The smallest available quant takes the same one-shot tests as Qwen. Two rounds of four tests, output checks, and a code error the model left behind.
1-bit to IQ4_XSFour quants
Qwen 3.8 Flash-Next: four quants on 128 GB
From 1-bit to IQ4_XS, four sizes take the same one-shot exam. Compare speed, output quality and the limits of each run.
QuantizationOne-shot tests
Qwen 3.8-27B: quantization and quality compared
From full BF16 to smaller quantized files and the default Ollama build. The same one-shot exam, with speed and quality side by side.
Model comparisonBF16
Qwen 3.8-27B vs Qwen 3.6-35B at BF16
Both models at full precision on the same machine. Four tests, their outputs, and the speed difference behind the quality scores.
Speed & reasoning
Change a setting. Measure the time, the output, and the tradeoff.
MTPBF16Repeated baseline
Qwen 3.8-27B: multi-token prediction speed tests
MTP off versus depths 2, 3 and 4 at BF16. Two protocols, a repeated baseline, and the memory and output tradeoffs behind the speedup.
Reasoning effortQuality vs time
Qwen 3.8-27B: reasoning levels compared
Off, low, medium and xhigh on the same tests. Compare quality and time, then inspect how the model's reasoning setting works.