llm-speed (2026-09-06). Gemma 4 12B IT QAT Q4_0 on RTX 5090: long-context measurements.
https://llm-speed.com/m/gemma-4-12b-it-qat#long-context

Figure: https://llm-speed.com/data/gemma-4-5090-context-2026-09-06-figure.png
Data: https://llm-speed.com/data/gemma-4-5090-context-2026-09-06.json
License: CC BY 4.0 — https://creativecommons.org/licenses/by/4.0/

Nine measurements: three sequential repetitions at each actual prompt length
(23,872 / 47,652 / 95,212 tokens), not estimated 32K/64K/128K targets.
Warm model, one active request, prompt caching disabled, thinking off.
llama.cpp CUDA revision 9725a313be0528214c4a02fed906ddaf7b3f712e.
Dots are medians; bars are observed minima and maxima, not confidence intervals.
No full 128K-input, maximum-context, multi-user throughput or answer-quality claim.

Source runs:
https://llm-speed.com/r/r_8q-uf-rq-gp
https://llm-speed.com/r/r_1srqmpjnkgd
https://llm-speed.com/r/r_-lh-vczf7ik

When adapting the figure, retain attribution and identify your changes.