Hardware limit
RAM, VRAM, CPU/GPU mode, host machine, container limit, and what normal desktop memory was already using.
Local LLM Test Records
Local LLM tests by RAM, CPU, and GPU.
This section keeps the local model test records: machine limits, runtime, settings, screenshots, measured speed, memory pressure, and where a model becomes too slow to use.
Last updated
2026-06-22
Current records
The page separates 8GB RAM CPU-only tests from 8GB VRAM GPU-fit tests. Each article keeps the setup, screenshots, raw notes, speed, and memory behavior close to the conclusion.
What every record keeps
A useful test record is more than a model name and a speed number.
The record is useful because it can be checked later: the machine, the settings, the raw proof, the measured result, and the point where the run stopped feeling practical.
Runtime and settings
Ollama or LM Studio version, model tag, quantization, context length, output limit, temperature, and install path.
Evidence kept
Screenshots, raw command output, Docker or nvidia-smi proof, model list, and test JSON when available.
Measured result
Tokens per second, wall time, loaded memory, peak GPU memory delta, and notes from the actual run.
Usability boundary
Where the model moved from comfortable to slow, tight, or not worth using as the daily default.
Next local check
The smaller or larger model a reader should try next on their own prompt before keeping a model.
Ollama model picks - 2026-06-16
Best Ollama Models for 8GB RAM
8GiB RAM CPU-only Ollama test record with install order, measured tokens per second, memory use, Docker screenshots, and model notes.
qwen2.5:0.5b firstgemma3:1b daily pickqwen2.5:7b is a slow boundary
CPU-only article - 2026-06-16
8GB RAM CPU-Only Local LLM Benchmark
8GiB Docker container, no GPU passthrough, six Ollama models, screenshots, memory notes, and measured tokens per second.
qwen2.5:0.5b: 58.02 tokens/sqwen2.5:7b: 7.15 tokens/sCPU-only, no GPU
GPU-fit article - 2026-06-17
8GB VRAM GPU Local LLM Fit Test
Ollama GPU test record with measured tokens per second, nvidia-smi screenshots, and peak GPU memory change under an 8GB VRAM budget.
qwen3:4b: 319.00 tokens/sqwen3.5:9b: 6.66GB peak deltagemma4:12b crosses 8GB
How to read these tests
Use the numbers as a machine check, not a speed guarantee
Step 1
Check the hardware limit first
Start with RAM, VRAM, CPU/GPU mode, runtime, and context settings before comparing model names.
Step 2
Compare fit and speed together
A model that loads is not automatically useful. Tokens per second, memory headroom, and later-turn slowdown all matter.
Step 3
Use the slow boundary
The slowest model that still runs is a warning line, not usually the model to recommend as the daily default.
Step 4
Retest your real prompt
Use the article to pick a nearby model, then rerun it on your own chat, coding, writing, or vision task.
Test record rule
Every test page must show the machine, settings, model list, speed, and slow boundary
Future articles can cover 16GB RAM, 8GB VRAM, 16GB VRAM, MacBook unified memory, and multi-GPU workstations without crowding the model finder page. The finder stays for broad matching; these pages are for readers who want real machine results before they install a model.
Questions
Local LLM benchmark FAQ
What is a local LLM real-world test?
It is a record from one fixed setup: machine or container limits, runtime, model list, settings, screenshots, memory behavior, and measured speed.
Should I trust tokens per second by itself?
No. Tokens per second is useful, but memory headroom, context length, answer quality, repeated turns, and whether normal desktop apps stay usable all matter.
Why separate CPU-only and GPU-fit tests?
CPU-only tests show the low-memory baseline, while GPU-fit tests show what changes when dedicated VRAM is available. Mixing them would make the result less useful.
Can I use these benchmarks for my exact laptop?
Use them as a starting point, not a speed promise. Your CPU, GPU, memory pressure, drivers, runtime, and prompt length can change the final result.
What should I test after reading a benchmark page?
Install the listed small model first, run one real prompt, watch memory pressure, then test a larger model only if the computer remains responsive.