Most people leave their local AI settings on defaults, but testing reveals that tweaking just a few parameters can dramatically improve how your local language model performs. One developer systematically tested the settings panel in LM Studio, a popular tool for running large language models (LLMs) locally on your own computer, and discovered that while the defaults work adequately, they're not optimized for individual use cases. The findings challenge the assumption that local AI configuration is either too complex or unnecessary to adjust.

Which Settings Actually Change Your Results?

When switching from cloud-based AI services to local models, users encounter a settings panel that can feel overwhelming. Unlike cloud AI where a company handles all the tuning behind the scenes, local AI puts control directly in your hands. The developer tested multiple parameters across different prompts and discovered that not all settings carry equal weight.

Temperature emerged as the most influential parameter. This setting controls how creative or predictable your model's responses become. Lower temperatures, around 0.3, produce more analytical and deterministic outputs, while higher temperatures closer to 1.0 generate more creative and varied responses. When tested with the same prompts at different temperature levels, the differences were stark. At 0.3, the model described "a shifting library," at 0.7 it became "a pruning garden," and at 1.0 it transformed into "a dimly lit room where the furniture moves when you're not looking." However, the impact varies significantly depending on the model size and type. Smaller, instruction-tuned models like Qwen 3.5 9b showed less variation across temperature settings than larger models would.

Presence penalty and repeat penalty both target redundancy but work differently. Repeat penalty operates at the token level, discouraging the model from reusing specific words and phrases. Presence penalty works more conceptually, pushing the model away from topics and ideas it has already covered, even when expressed in different words. Testing revealed that repeat penalty functions primarily as a stability control rather than a quality improvement tool. Pushing it above 1.1 on smaller models produced unexpectedly long and tedious responses. In contrast, presence penalty offers more tuning flexibility, with nudging it between 0.7 and 1.0 producing the most coherent yet varied writing for general use.

What About the Settings Everyone Recommends?

Two widely recommended parameters, Top-K and Top-P, turned out to be less critical than many assume. These settings use fixed rules to filter which tokens the model can choose from next. Min-P, by comparison, scales dynamically based on what the model is actually doing in that moment. Min-P sets a floor on token probability, cutting out tokens that aren't competitive relative to the model's best guess while preserving the creative energy from higher temperatures. This dynamic approach explains why most people in the local LLM community have converged on temperature and Min-P as the primary settings worth adjusting, leaving Top-K and Top-P at their defaults.

Min-P prevents the incoherence that can result from high temperatures. When you increase temperature to get more creative outputs, the token probability distribution flattens, making nonsensical tokens as likely as good ones. Min-P acts as a leash, letting the model run freely while preventing it from veering into gibberish. Testing showed that pushing Min-P too high makes the model overly cautious, replicating the problem of low temperature from a different angle. The sweet spot for general use fell around 0.1, where short story generation became noticeably tighter and more coherent compared to a Min-P setting of 0.

How to Optimize Your Local LLM Settings

• Start with Temperature: Begin by adjusting temperature if you only change one setting. Push it higher for creative tasks and lower for analytical work. Test the same prompt at 0.3, 0.7, and 1.0 to see how your specific model responds.
• Fine-tune Presence Penalty: Once temperature is set, adjust presence penalty between 0.7 and 1.0 to prevent the model from circling back to the same ideas repeatedly. This handles conceptual redundancy rather than just repeated words.
• Use Min-P as a Safety Valve: If high temperature produces incoherent outputs, increase Min-P gradually starting from 0 to find the balance point where creativity remains but gibberish disappears.
• Keep Repeat Penalty Stable: Leave repeat penalty at 1.0 for most use cases. Only adjust it if you notice excessive word repetition, but avoid pushing it above 1.1 on smaller models.
• Test One Variable at a Time: Change a single parameter, run the same prompt multiple times, and observe the results before adjusting anything else. This prevents confusion about which setting caused which change.

The developer emphasized that no universal prescription exists for these settings. The same temperature value behaves differently depending on the model's size, quantization level, and how other parameters interact. A 9-billion parameter model like Qwen 3.5 9b will respond differently to the same adjustments than a 14-billion parameter model. This variability is precisely why local AI offers value that cloud services cannot: the ability to tune your setup for your specific needs and hardware.

The defaults in LM Studio are not bad, but they represent a starting point rather than an optimized endpoint. Users who invest time in understanding how temperature shapes personality, how presence penalty prevents laziness, and how Min-P prevents high-temperature chaos gain meaningful control over their local AI experience. This hands-on tuning capability distinguishes local models from cloud-based alternatives where such customization is impossible.