Is unsloth/gemma-4-12B-it-qat-GGUF good for coding?

In our benchmarks, unsloth/gemma-4-12B-it-qat-GGUF scores 60.3/100 for coding. It runs at about 61.2 tok/s, so if you want more speed, Ornith-1.5-35B-A3B.APEX-I-NanoPlus is faster (~110.5 tok/s) and still scores well for coding (81.1/100).

Is unsloth/gemma-4-12B-it-qat-GGUF good for agentic (tool-using) tasks?

In our benchmarks, unsloth/gemma-4-12B-it-qat-GGUF scores 68.8/100 for agentic workflows. It runs at about 61.2 tok/s, so if you want more speed, swift-iq2_xs is faster (~91.5 tok/s) and still scores well for agentic workflows (91.4/100).

How fast is unsloth/gemma-4-12B-it-qat-GGUF for local inference?

Across 7 community benchmark runs, unsloth/gemma-4-12B-it-qat-GGUF reaches up to 82.0 tok/s and averages 61.2 tok/s, with the fastest results on NVIDIA GeForce RTX 4070.

How much memory does unsloth/gemma-4-12B-it-qat-GGUF need?

The leanest observed configuration used about 10.3 GB of memory (quantizations tested: UD-Q4_K_XL).

Which tools have been used to run unsloth/gemma-4-12B-it-qat-GGUF?

Benchmarks were submitted using Unsloth Studio, llama.cpp. Results are community-contributed and updated as new runs arrive.

Model benchmarks

unsloth/gemma-4-12B-it-qat-GGUF local LLM performance

As of October 2026, unsloth/gemma-4-12B-it-qat-GGUF runs at up to 82.0 tok/s for local inference (best of 7 community benchmark runs across 2 GPUs).

Unsloth Studiollama.cppUD-Q4_K_XL

ShareReddit X

Model size

12B

Peak speed

82.0 tok/s

Average speed

61.2 tok/s

Avg PP

950.1 tok/s

Min memory

10.3 GB

Max context

100,000 tokens

Avg output / run

11,533 tokens

Avg runtime / run

3m 30s

Avg quality

70.4

Benchmark runs

GPUs tested

Quality by task

Average LLM-judged quality (0–100) with the run-to-run spread shown as a P5–P95 band, overall and for each benchmark task, across all 7 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.

Task	P5 (low)	Avg	P95 (high)
Overall	67.9	70.4	73.3
Agent Workflow	59.2	68.8	77.1
Code Generation	47.6	60.3	71.5
Role Play & Narrative	79.9	81.9	84.8
Research & Analysis	66.1	70.8	75.6

Performance by hardware and tool

Every hardware/tool/quantization combination unsloth/gemma-4-12B-it-qat-GGUF has been benchmarked on, ranked by peak token generation speed. Last updated October 2026.

Hardware	Tool	Quant	Best tok/s	Avg tok/s	Memory	Context	Quality	Runs
NVIDIA GeForce RTX 4070	Unsloth Studio	UD-Q4_K_XL	82.0 tok/s	64.9 tok/s	10.3 GB	100,000 tokens	71.9	4
AMD Radeon RX 7900 GRE	Unsloth Studio	UD-Q4_K_XL	67.2 tok/s	67.2 tok/s	11.5 GB	65,536 tokens	68.1	1
NVIDIA GeForce RTX 4070	llama.cpp	—	51.4 tok/s	50.7 tok/s	n/a	32,768 tokens	68.7	2

Benchmark runs

All 7 unsloth/gemma-4-12B-it-qat-GGUF runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.

Open run page →

Frequently asked questions

Is unsloth/gemma-4-12B-it-qat-GGUF good for coding?: In our benchmarks, unsloth/gemma-4-12B-it-qat-GGUF scores 60.3/100 for coding. It runs at about 61.2 tok/s, so if you want more speed, Ornith-1.5-35B-A3B.APEX-I-NanoPlus is faster (~110.5 tok/s) and still scores well for coding (81.1/100).
Is unsloth/gemma-4-12B-it-qat-GGUF good for agentic (tool-using) tasks?: In our benchmarks, unsloth/gemma-4-12B-it-qat-GGUF scores 68.8/100 for agentic workflows. It runs at about 61.2 tok/s, so if you want more speed, swift-iq2_xs is faster (~91.5 tok/s) and still scores well for agentic workflows (91.4/100).
How fast is unsloth/gemma-4-12B-it-qat-GGUF for local inference?: Across 7 community benchmark runs, unsloth/gemma-4-12B-it-qat-GGUF reaches up to 82.0 tok/s and averages 61.2 tok/s, with the fastest results on NVIDIA GeForce RTX 4070.
How much memory does unsloth/gemma-4-12B-it-qat-GGUF need?: The leanest observed configuration used about 10.3 GB of memory (quantizations tested: UD-Q4_K_XL).
Which tools have been used to run unsloth/gemma-4-12B-it-qat-GGUF?: Benchmarks were submitted using Unsloth Studio, llama.cpp. Results are community-contributed and updated as new runs arrive.