7 parameters that control every LLM response explained
blaizedsouza · x · 2026-08-20
This post breaks down the seven key parameters that dictate how LLMs sample from their probability distribution:
- Max tokens: A hard cap on output length. Low values truncate sentences rather than yielding concise answers.
- Temperature: Scales scores to adjust probabilities. <1 sharpens the distribution towards likely tokens, >1 flattens it for more randomness, and 0 enables deterministic output.
- Topp: Retains the smallest set of tokens whose cumulative probability reaches p, sampling only from that group to adapt to model confidence.
The linked article also provides a first-principles tour of inference mechanics, including tokenization, embeddings, attention, prefill/decode split, KV caching, and quantization.
More from Models
- Tencent Starts Gray Testing Flagship Hunyuan Hy4 Model — Nunki08 · 2026-08-20
- Mini Kimi-K3 Replicated Under $250 Beats GPT-2 Benchmark — OtherRaisin3426 · 2026-08-20
- User drains Codex quota in hours using only Luna Max — Al_Grigor · 2026-08-20
- Qwen 2.5 72B Aces ACT Exam with Perfect Reading Score — on_line187 · 2026-08-20
- Claude Advances Math Bound While Attempting Riemann Hypothesis — PtrPomorski · 2026-08-20
- LLMs Struggle with Simple Anagram Task: Claude Succeeds, Others Fail — HydronautInSpace · 2026-08-20