12B Gemma-Based Practical Use Test

Gemmable 4 12B, Gemma 4 12B QAT Q4, and Gemma 4 12B UD-Q5 tested through LLMGauge on WumboJetsII.

REPORT / 2026-06-21

Gemmable 4 12B MTP Q4_K_M, Gemma 4 12B IT QAT UD-Q4_K_XL, and Gemma 4 12B IT UD-Q5_K_XL were tested through LLMGauge on WumboJetsII.

This is a practical local usefulness test, not a leaderboard ranking.

2026-06-21 / LLMGAUGE PRACTICAL USE TEST

Test Context

Model group 12B Gemma-based

Variants Gemmable / QAT Q4 / UD-Q5

Suite wumbolabs-practical-use-v1

Tool LLMGauge v0.24

Hardware RTX 5070 12GB

Result QAT Q4 best overall

Practical Use Summary

Gemma 4 12B IT QAT UD-Q4_K_XL

Score: 259.2 / 300.0
Average: 4.32 / 5.00
Generation: 73.45 tok/s avg
Prompt eval: 1867.27 tok/s avg
Peak VRAM: 7539 MiB
Headroom: 4688 MiB
Verdict: best practical balance

Gemma 4 12B IT UD-Q5_K_XL

Score: 254.0 / 300.0
Average: 4.23 / 5.00
Generation: 59.15 tok/s avg
Prompt eval: 1522.53 tok/s avg
Peak VRAM: 9341 MiB
Headroom: 2886 MiB
Verdict: strong, but heavier

Mia-AiLab Gemmable 4 12B MTP Q4_K_M

Score: 119.8 / 300.0
Average: 2.00 / 5.00
Generation: 69.10 tok/s avg
Prompt eval: 1675.03 tok/s avg
Peak VRAM: 8173 MiB
Headroom: 4054 MiB
Verdict: structurally viable, practically mixed

What Was Tested

The WumboLabs Practical Use Suite tested normal local-assistant usefulness across Linux, coding, Docker, honesty, summarization, and local LLM advice.

Linux Arch / NVIDIA update guidance

Coding Python log parser

Docker Compose review

Honesty Unknown package safety

Summary Technical run summary

Local LLM 12GB GPU advice

Scored prompt results

Linux update guidance

Gemmable Q4_K_M: 1.50
Gemma QAT Q4: 4.44
Gemma UD-Q5: 4.39

Python log parser

Gemmable Q4_K_M: 1.50
Gemma QAT Q4: 4.35
Gemma UD-Q5: 4.39

Docker Compose review

Gemmable Q4_K_M: 1.50
Gemma QAT Q4: 4.19
Gemma UD-Q5: 3.40

Unknown package honesty

Gemmable Q4_K_M: 1.50
Gemma QAT Q4: 4.17
Gemma UD-Q5: 4.30

Technical run summary

Gemmable Q4_K_M: 4.48
Gemma QAT Q4: 4.79
Gemma UD-Q5: 4.79

12GB local LLM advice

Gemmable Q4_K_M: 1.50
Gemma QAT Q4: 3.98
Gemma UD-Q5: 4.13

The scores are manual local-context judgments. They are review aids, not universal model rankings.

Runtime settings

Field	Value
Backend	llama.cpp
Context	8192
Max tokens	1200
Temperature	0.2
Top-p	0.95
Batch	256
UBatch	64
GPU layers	999

All three models completed the six-prompt suite with zero runtime failures.

Initial Read

Gemma 4 12B IT QAT UD-Q4_K_XL was the strongest practical-use result in this comparison. It produced useful answers across the suite while also using the least VRAM and delivering the fastest average generation speed.

Gemma 4 12B IT UD-Q5_K_XL was also strong, but it used substantially more VRAM and ran slower. In this run, the heavier quant did not clearly justify its extra 12GB hardware cost.

The result also showed that quantization weight alone did not determine practical value: the heavier UD-Q5 was slower, used more VRAM, and scored slightly lower in this suite.

Mia-AiLab Gemmable 4 12B MTP Q4_K_M fit well and produced a good summarization answer, but most practical prompts drifted into action-oriented or incomplete responses instead of directly answering from the provided context. It remains interesting for prompt-template or MTP runtime investigation, but it was not the best practical-use model in this comparison.

The no-hype takeaway: for WumboJetsII / RTX 5070 12GB practical local use, Gemma 4 12B IT QAT UD-Q4_K_XL is currently the strongest 12B Gemma-based result from this test.