Maximizing Model Performance Guide

Maximizing Model Performance Guide

Guides you in tuning samplers and parameters for local language models, with steps for system prompts, reasoning models and expert settings in LM Studio, SillyTavern, Ollama and Koboldcpp.

Made by DavidAU on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a guide from DavidAU for getting better results out of local language models by tuning samplers, advanced samplers and parameters. It covers every quantization type and full precision. The part shown links four extra documents, and the main guide continues past what is visible here.

Who it suits: People who run language models on their own computer and want to tune how those models behave. It suits users of LM Studio, SillyTavern, Ollama and Koboldcpp, and anyone running DavidAU's own models.

Industry | Who uses it | How they use it | Example
Local AI | Hobbyists | Tune how a local model behaves | Paste a system prompt into the app to switch reasoning on
Role play | SillyTavern users | Set the system prompt on the template page | Activate the system prompt option and enter the text
Development | Developers | Set how many experts a mixture-of-experts model uses | Follow the guide's steps for the app you run
Model tinkering | Model tinkerers | Turn a regular model into a reasoning one | Follow the adapter steps that use Mergekit

You can skip it if you only use hosted chat services, since the guide is about local apps.

Where to use it:

• LM Studio
• SillyTavern
• Ollama
• Koboldcpp

Use case | Who | What they get back
Reasoning control | Local model users | Reasoning switched on or off through the system prompt
Expert settings | Developers | A chosen number of active experts in a mixture-of-experts model
Low quant support | Users of small quants | Software the author says helps his models run without issues

How to set it up: You need Python for the download tool and a local app to apply the settings in.

• Python, to run the Hugging Face tool
• A local app such as LM Studio, SillyTavern, Ollama or Koboldcpp
• A model to tune, such as one of the author's

1. Install Python if you do not have it.
2. Install the Hugging Face hub tool.
3. Download the repository.
4. Open your app and find its system prompt setting.

pip install -U huggingface_hub
hf download DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters

How to use it: Open your app's system prompt box and paste the text a model's page gives you. Keep every line break when you paste. If you need to edit the text, do it in Notepad or in the app itself.

Tool | What it does
LM Studio | Switch on Power User or Developer mode, then paste into the system prompt box
SillyTavern | On the template page, activate the system prompt and enter the text
Ollama | Set the system message, as Ollama's own readme describes
Koboldcpp | Load and start the model, open settings, choose a template and fill the system prompt box

Example: You load a reasoning model in LM Studio and turn on Developer mode. You paste the author's reasoning system prompt into the box, keeping the line breaks. The model now writes out its thinking. With the box empty, a model that needs the prompt acts like a normal chat model.

For newcomers and experts: Samplers are settings that control how a model picks its next words. A system prompt is standing text that tells the model how to behave. A first small step is to open your app's system prompt box and see what is there.

The guide calls the system prompt root access to the model. It steers instruction following, output and, for reasoning models, whether reasoning is on. A larger model, or one that follows instructions well, reacts more to a pasted prompt. The linked documents also cover adapters, a slice of a reasoning model, that turn a regular model into a thinking one.

Pick a local app and a model
Open the app's system prompt setting
Paste the prompt text with line breaks
Run the model and compare the output

Good to know:

• The main sampler and parameter guide is cut off in the part shown, so this page covers the extra documents and system prompt steps.
• For reasoning models that need it, no system prompt means reasoning stays off.
• The card calls the system prompt a way to also improve ordinary models, with a stronger effect on larger ones.
• License: apache-2.0.
• Cost to run: none named, you use your own hardware.
• Data it reaches: files downloaded from Hugging Face, plus the pages its links lead to.
• Not for: chatting with directly, since the card reads as a guide rather than a model.

On Hugging Face, 13 November 2026

• Downloads in the last 30 days: 0
• Likes: 276
• License: apache-2.0
• Tags: coherence, llamacpp, lmstudio, sillytavern, koboldcpp, backyard, ollama, steering, ggufs, exl2, quants, imatrix

Links

• Model: https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters
• Files: https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters

Then follow the usage in its model card. Nothing in this kit runs by itself.