Gemma 4 31B

google/gemma-4-31B-it

30.7BDenseText generationImage → textText inputImage input

Google · 2 recorded configurations

Measured setup

Hardware

1× RTX 3090 · GPU-resident
2× RTX 3090 · GPU-residentalso Runtime

Runtime and tool

live llama-server
llama-benchalso Hardware

Quantisation

Q4_K_M

MTP draft depth

MTP off

Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.

PP · left axisTG · right axis

View exact measurements

Exact context depth and measured fresh prefill, bulk prefix, and decode rates for Gemma 4 31B QAT
Context depth (tokens)	Fresh prefill PP (tok/s)	TG (tok/s)
512	1,087.93	34.37
2,048	1,076.66	34.31
4,096	1,050.22	34.28
8,192	1,017.71	34.24
16,384	943.72	34.19
32,768	836.60	34.14

Technical record and provenance

Run details

Source run ID: rtx3090x1-gemma-4-31b-q4km-2026-07
Hardware: 1× RTX 3090
Runtime: llamaCpp
Weight placement: gpu-resident
KV-cache precision: Q8_0
Prompt tokens: 512
Generated tokens: 128
Speculative decoding: None (standard decoding)
Notes: Single RTX 3090, llama.cpp, Q4_K_M, Q8_0 KV cache. Variant tested is the QAT uncensored finetune by HauhauCS. Build, driver, power limit, and repetitions not recorded in the article.
Source article: /local-ai/rtx-3090-for-local-ai-inference/

Compare measured entries