Is google/gemma-4-12b good for coding?

In our benchmarks, google/gemma-4-12b scores 62.2/100 for coding. It runs at about 40.3 tok/s, so if you want more speed, nvidia/Qwen3.8-27B-NVFP4 is faster (~63.3 tok/s) and still scores well for coding (82.7/100).

Is google/gemma-4-12b good for agentic (tool-using) tasks?

In our benchmarks, google/gemma-4-12b scores 78.6/100 for agentic workflows. It runs at about 40.3 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 google/gemma-4-12b for local inference?

Across 7 community benchmark runs, google/gemma-4-12b reaches up to 91.7 tok/s and averages 40.3 tok/s, with the fastest results on NVIDIA GeForce RTX 5080.

How much memory does google/gemma-4-12b need?

The leanest observed configuration used about 9.2 GB of memory.

Which tools have been used to run google/gemma-4-12b?

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

Model benchmarks

google/gemma-4-12b local LLM performance

As of October 2026, google/gemma-4-12b runs at up to 91.7 tok/s for local inference (best of 7 community benchmark runs across 5 GPUs).

LM Studiollama.cpp

ShareReddit X

Model size

12B

Peak speed

91.7 tok/s

Average speed

40.3 tok/s

Avg PP

516.0 tok/s

Min memory

9.2 GB

Max context

131,072 tokens

Avg output / run

10,575 tokens

Avg runtime / run

10m 59s

Avg quality

72.1

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	66.5	72.1	74.7
Agent Workflow	70.9	78.6	83.6
Code Generation	54.1	62.2	67.4
Role Play & Narrative	59.2	76.6	88.3
Research & Analysis	61.8	71.1	78.1

Performance by hardware and tool

Every hardware/tool/quantization combination google/gemma-4-12b 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 5080	llama.cpp	—	91.7 tok/s	91.1 tok/s	9.2 GB	131,072 tokens	74.3	2
AMD Radeon RX 6800 XT	LM Studio	—	32.4 tok/s	32.4 tok/s	n/a	8,192 tokens	73.4	1
NVIDIA GeForce RTX 4070 Ti SUPER	LM Studio	—	24.9 tok/s	24.9 tok/s	n/a	16,384 tokens	73.3	1
CPU only	LM Studio	—	21.6 tok/s	21.6 tok/s	n/a	8,192 tokens	73.1	1
Intel(R) Arc(TM) Pro 140T GPU (16GB)	LM Studio	—	10.8 tok/s	10.8 tok/s	n/a	8,192 tokens	63.7	1
AMD Radeon RX 6600	LM Studio	—	10.0 tok/s	10.0 tok/s	n/a	8,192 tokens	73.1	1

Benchmark runs

All 7 google/gemma-4-12b runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.

Open run page →

Frequently asked questions

Is google/gemma-4-12b good for coding?: In our benchmarks, google/gemma-4-12b scores 62.2/100 for coding. It runs at about 40.3 tok/s, so if you want more speed, nvidia/Qwen3.8-27B-NVFP4 is faster (~63.3 tok/s) and still scores well for coding (82.7/100).
Is google/gemma-4-12b good for agentic (tool-using) tasks?: In our benchmarks, google/gemma-4-12b scores 78.6/100 for agentic workflows. It runs at about 40.3 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 google/gemma-4-12b for local inference?: Across 7 community benchmark runs, google/gemma-4-12b reaches up to 91.7 tok/s and averages 40.3 tok/s, with the fastest results on NVIDIA GeForce RTX 5080.
How much memory does google/gemma-4-12b need?: The leanest observed configuration used about 9.2 GB of memory.
Which tools have been used to run google/gemma-4-12b?: Benchmarks were submitted using LM Studio, llama.cpp. Results are community-contributed and updated as new runs arrive.