How much VRAM does gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic need?

Q4_K_M is exactly 7,381,386,496 bytes (6.87 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

llmfan46 · text

gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic

llmfan46/gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic

gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic at Q4_K_M is exactly 7,381,386,496 bytes (6.87 GiB / 7.38 GB) — an effective 4.938 bits per weight, not the nominal 4. Its KV cache at 32K is 2.47 GiB, not the 12.00 GiB a flat formula predicts.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

12.0B

Architecture

gemma4

48 layers

Context

262,144

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q2_K	4.50 GiB	4,830,151,936	3.231	—	mradermacher
Q3_K_S	5.15 GiB	5,528,233,216	3.698	—	mradermacher
Q3_K_M	5.67 GiB	6,087,091,456	4.072	—	mradermacher
Q3_K_L	6.12 GiB	6,566,323,456	4.392	—	mradermacher
IQ4_XS	6.23 GiB	6,690,555,136	4.475	—	mradermacher
Q4_K_S	6.54 GiB	7,024,051,456	4.699	—	mradermacher
Q4_K_M	6.87 GiB	7,381,386,496	4.938	—	mradermacher
Q4_0	7.07 GiB	7,596,302,496	5.081	—	llmfan46
NVFP4	7.60 GiB	8,163,421,312	5.461	—	llmfan46

KV cache by context

computed per layer — this model uses sliding-window attention

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.72 GiB	1.50 GiB	2.09×	8 / 40 / 0
8,192	0.97 GiB	3.00 GiB	3.10×	8 / 40 / 0
16,384	1.47 GiB	6.00 GiB	4.09×	8 / 40 / 0
32,768	2.47 GiB	12.00 GiB	4.86×	8 / 40 / 0
65,536	4.47 GiB	24.00 GiB	5.37×	8 / 40 / 0
131,072	8.47 GiB	48.00 GiB	5.67×	8 / 40 / 0

40 of 48 layers cache only a 1,024-token window rather than the full context, on a period of . Figures assume the default configuration; --swa-full disables the saving entirely.

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-27B vs Qwen3.8-27B vs gemma-4-12B-it-qat-q4_0-unquantized vs gemma-4-E4B-it vs Qwen3-30B-A3B-Thinking-2507 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 6.27 GiB. The real file is 6.87 GiB, because a quantization is a mixture and some tensors are always kept at higher precision. The larger discrepancy is the cache: a flat formula gives 12.00 GiB at 32K context where the real figure is 2.47 GiB, because most of this model's layers cache a fixed window rather than the whole context.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

256

Hidden size

3840

Vocab

262,144

Sliding window

1024

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic need?: Q4_K_M is exactly 7,381,386,496 bytes (6.87 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic's KV cache?: 2.47 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Which quantization of gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.