The longer I use my local LLMs, the more I realize how much of the "local models feel worse" experience is often just a settings problem. I've spent the last few months playing around with a bunch of settings and parameters. Local AI is still a relatively new concept to me and I'm learning as I go, so naturally, I often come across useful tips that should have been so obvious to me in hindsight.

I'm using LM Studio as my runner, primarily because of the easy interface. It comes with a bunch of bits and bobs that let you customize your session experience, and one I've been playing with a lot lately is presets. I knew they existed but never fully took advantage of them until I expanded my use of local AI - from private health stuff to learning materials and storytelling. That's when I realized I needed a quicker way to tune my model's parameters when flipping between sessions.

The defaults will fail you

Your tuned setup doesn't survive the session

When you open a new chat in LM Studio without any saved configurations, you'll be met by default parameters in the sidebar. LM Studio doesn't have app-wide parameter defaults, but every model does have model-wide parameter defaults, which you can change if you want to (go to My Models > Settings icon > Inference tab). Thing is, even if you change these model defaults to something else, every new session thereafter will still snap back to the model's defaults again.

The system prompt field is the bigger one though - that opens blank every time unless you also add a system prompt to the model defaults. So unless you configured your model, whatever instructions you had last session, whatever persona or format rules or conciseness constraints - none of it carries.

Adding a system prompt to your model and configuring its sliders can be useful if you tend to reach for the same model for tasks that fall under the same category. But that wasn't the case for me. I used to stick to gpt-oss 20B religiously because it's one of the top general-purpose models. Now my go-to is Qwen 3.5 9b, which I use for pretty much anything regardless of the topic - learning guidance, stories, wellness information, just a wide range of topics.

This range calls for various parameter configurations. There was something in LM Studio for this exact scenario, staring me in the face the whole time.

This is what presets are for

They prep your chat so you don't have to

Presets are basically named configuration profiles for your local AI sessions. The concept isn't LM Studio-specific - other runners have their own versions, Ollama does it through Modelfiles, for example. But LM Studio's implementation is the most approachable if you're working through a GUI rather than a terminal.

A preset bundles together a system prompt and every parameter in the Advanced Configuration sidebar into a single package you can save, name, and load into any chat in one click. It sounds simple, because it is, but this little package packs so much power. Temperature, presence penalty, min-p, system prompt, and all of it stored together under whatever you want to call it. They're not model-specific either, so the same preset works across different models, and overrides the model's defaults.

Setting one up takes about thirty seconds. When you've been chatting and tweaking the sliders, and chatting some more to test the outcomes again, and finally found your sweet spot, that's when I recommend creating the preset. You'll see Presets at the top of the Configurations tab, and there will be a dropdown button labeled "unsaved preset". Hit that, then New Preset, give it a name, hit save, and hit save again.

That's all there is to it. Now you can find that preset in the dropdown in any session, regardless of your model. They're stored as plain JSON files locally too, so if you want to edit one directly or back them up, you can do that in any text editor.

What a workflow with presets looks like

And the benefits they bring

I have a bunch of presets, but only use three of them regularly. One for coursework - tighter temperature, conciseness instruction in the system prompt, presence penalty nudged up to stop Qwen from circling the same point four different ways. One for general back-and-forth, which is honestly just the model defaults with a system prompt that instructs it to use Brave Search MCP to fill in its knowledge gaps. And then one for helping me with brainstorming ideas for my novel - so temperature high for more creativity and a slightly lower penalty number.

The less obvious thing is that presets work as a record of your own configuration decisions. If you've tested your way to a temperature and penalty combo that actually works for a specific use case, you're not relying on memory to reconstruct it next time, and you're actually getting the responses you need.

And as of LM Studio 0.3.15, you can publish presets to the LM Studio Hub and import ones other people have built - so if you're still figuring out where to even start with parameters, that's a reasonable shortcut.

One dropdown away from a better session

The defaults in LM Studio aren't bad, they're just not mine. Presets are what close that gap - one setup for documents, one for creative work, one for whatever else I reach for the model for. And once your configurations are saved, you stop second-guessing whether the output feels off because of your prompt or because you forgot to re-tune the sidebar again.