Model parameters (often called valves in the interface) tune how a model generates answers. They appear in chat controls, templates, custom models and personal defaults. Your exact list depends on the selected model.
The main five
Temperature
How adventurous the word choices are.
| Value | Character | Use for |
|---|---|---|
| 0.1–0.3 | Deterministic, consistent | Extraction, classification, factual Q&A, code |
| 0.4–0.7 | Balanced | General business writing, summaries |
| 0.8–1.2 | Creative, varied | Brainstorming, marketing copy, fiction |
Rule of thumb: anything that must be right runs low; anything that must be fresh runs high. Change temperature or top-p, not both.
Top-P (nucleus sampling)
Alternative randomness control — considers the top P of probability mass. 0.9 is a common default; lower = tighter. Adjust as an alternative to temperature.
Max tokens
The ceiling on answer length (≈ 0.75 words per token in English). Set to control verbosity and cost; leave unset for the model's natural limit.
Frequency & Presence penalty
Discourage repeating the same words (frequency) or topics (presence). Useful when a model starts echoing. Small values (0.1–0.5) are enough; high values degrade quality.
Stop
Character sequences that end generation immediately — handy for structured formats.
Reproducibility and context
| Parameter | Meaning | Notes |
|---|---|---|
| Seed | Fixed randomness start. Same seed + same input ≈ same output. | Fix it when reproducing a good answer; randomise for variety. |
| Context length | Working memory of the conversation, in tokens. | Raising it increases memory use and latency; if answers start "forgetting", check it first. |
| Keep Alive | How long a model stays loaded in memory after use. | Higher = faster first reply, more reserved GPU. Busy shared systems leave this short. |
| Num batch / threads | Internal throughput controls for local models. | Defaults are tuned; adjust only with your admin. |
Reasoning display
| Parameter | Effect |
|---|---|
| Reasoning / Show thinking | Displays the model's intermediate reasoning where the model supports it. Great for auditing; costs extra tokens. |
Recommended starting points
| Task | Settings |
|---|---|
| Document extraction/classification | Temperature 0.1, top-p default |
| Factual Q&A over knowledge bases | Temperature 0.2 |
| Emails and business documents | Temperature 0.5 |
| Brainstorming/marketing copy | Temperature 0.9 |
| Code generation | Temperature 0.2–0.4 |
| Debugging/explaining code | Temperature 0.4–0.6 |
Note:
Parameters change behaviour, not knowledge. If a model lacks your facts, fix it with a knowledge base or fine-tuning — no parameter substitutes for the right information.