Gemini last models: temperature, top_p, and top_k are deprecated and ignored

**temperature**, **top_p**, and **top_k** are deprecated and ignored on Gemini’s latest models. The change surfaced on Hacker News’s front page, where the discussion focused on what that means for people who still pass those sampling controls into Gemini APIs or SDKs.

Those three knobs control how the model picks the next token. **temperature** scales the logits before sampling, so higher values spread probability mass and lower values sharpen it. **top_k** keeps only the k highest-probability tokens; **top_p** (nucleus sampling) keeps the smallest set whose cumulative probability reaches p. On the affected Gemini models, callers can still send the fields, but the service does not apply them—the generation path runs without that sampling stage.

For engineers and builders, any stack that tunes creativity, determinism, or style through those parameters will not behave as configured on those models. Prompt templates, eval harnesses, and A/B setups that vary temperature or nucleus settings will look like they still accept the options, yet outputs will not track those knobs. That breaks assumptions in chat UIs, agent loops, content pipelines, and regression suites that treat sampling config as a reliable control surface.

In the wider market, most major LLM APIs still document and honor temperature and nucleus-style controls as first-class generation settings. Gemini deprecating and ignoring them on its latest models is a product divergence: less user-facing sampling control, more fixed server-side decoding. That matters when you multi-homed prompts across providers or assumed a shared “knob language” for quality and diversity.

If you call Gemini’s latest models, stop treating **temperature**, **top_p**, and **top_k** as live levers—remove them from critical paths or gate them so they only apply on models that still honor them. Watch for official guidance on which models ignore the parameters, whether any replacement controls appear, and whether older models keep the old behavior so you can pin versions where sampling still matters.