Leaderboard/gemma-4-12b-it-qat/RTX 5090 (32GB)/r_-lh-vczf7ik

gemma-4-12b-it-qat on RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB

Name: gemma-4-12b-it-qat on RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB
Creator: llm-speed
License: https://creativecommons.org/licenses/by/4.0/
Keywords: LLM benchmark, tokens per second, gemma-4-12b-it-qat, RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB, llama.cpp

RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB

suite suite-v1

cli 0.0.5

signed0tv44ISLy1…

Embed badgesubmitted Sep 06, 2026

https://llm-speed.com/r/r_-lh-vczf7ik

Workload results

Workload	Backend	Model	decode tok/s	prefill tok/s	TTFT	p50	p95
long-context-decay	llama.cpp@1 (9725a31)	gemma-4-12b-it-qatQ4_0	142.6tok/s	7,115.8tok/s	3,355ms	7.0ms	7.0ms

Long-context results

gemma-4-12b-it-qat · llama.cpp. The workload overview shows the smallest successful point. This table exposes every reported target, including unsuccessful or missing measurements.

Targets estimate tokens from character count. Actual prompt tokens come from the runtime; neither is the configured context capacity. “—” means unmeasured or unavailable.

Scroll the table sideways for all measurements.

Long-context measurements and status for each estimated target.
Estimated target	Actual prompt tokens	First token (s)	Decode tok/s	Output tokens	Status
32,000	23,872	3.355	142.61	98	measured
64,000	47,652	8.833	133.03	100	measured
128,000	95,212	20.224	123.65	96	measured

First-token delay includes processing the prompt with the model loaded. These speed measurements do not test answer quality or the maximum usable context.

Part of our Gemma 4 RTX 5090 long-context study: see all three repetitions, observed ranges and the 131,072-context server configuration.

Reproduce on your machine

Start with this model and workload selection. This command does not capture the original artifact, backend version or server settings:

$ pipx install llm-speed && llm-speed bench --model 'gemma-4-12b-it-qat' --workload 'long-context-decay'

Match quantization, context capacity, caching and runtime settings for a useful comparison. Duration varies by workload and hardware; the published client may differ from this run's version. How it's measured.

Embed this run

Drop the badge into a README, blog post, or signature. Each render is a backlink to the signed result.

[![llm-speed: 143 tok/s on RTX 5090 (32GB) (gemma-4-12b-it-qat)](https://llm-speed.com/badge/r_-lh-vczf7ik.svg)](https://llm-speed.com/r/r_-lh-vczf7ik)

Related benchmarks

More gemma-4-12b-it-qat benchmarks — every backend and rig that has run this model.
More RTX 5090 (32GB) LLM benchmarks — every model measured on this hardware.

Provenance

Run ID: r_-lh-vczf7ik
Fingerprint hash: 013ca61a09d17996
Public key: 0tv44ISLy10gz6Oc6FJig3eWJVLwso9oKemSM4nicKM=
Received: 2026-09-06 03:40:00