How much VRAM does gemma-4-A4B-98e-v7-coderx-it need?

Q4_K_M is exactly 13,239,744,800 bytes (12.33 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is gemma-4-A4B-98e-v7-coderx-it's KV cache?

1.54 GiB at 32K context with an f16 cache, computed per layer from the model's architecture — most of its layers use sliding-window attention, so a flat formula overstates this substantially. Quantizing the cache to q8_0 roughly halves it.

Is gemma-4-A4B-98e-v7-coderx-it a mixture-of-experts model?

Yes — 98 experts with null routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

ManniX-ITA · text · mixture of experts

gemma-4-A4B-98e-v7-coderx-it

ManniX-ITA/gemma-4-A4B-98e-v7-coderx-it

gemma-4-A4B-98e-v7-coderx-it at Q4_K_M is exactly 13,239,744,800 bytes (12.33 GiB / 13.24 GB) — an effective 5.179 bits per weight, not the nominal 4. Its KV cache at 32K is 1.54 GiB, not the 7.50 GiB a flat formula predicts.

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

Parameters

20.5B

total, not active

Architecture

gemma4

30 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	8.22 GiB	8,822,091,552	3.451	—	ManniX-ITA
Q3_K_M	9.79 GiB	10,512,916,512	4.112	—	ManniX-ITA
Q3_K_L	10.19 GiB	10,942,874,656	4.281	—	ManniX-ITA
IQ4_XS	10.25 GiB	11,008,234,016	4.306	—	ManniX-ITA
IQ4_NL	10.63 GiB	11,418,356,256	4.467	—	ManniX-ITA
Q4_K_S	11.37 GiB	12,209,640,736	4.776	—	ManniX-ITA
Q4_K_M	12.33 GiB	13,239,744,800	5.179	—	ManniX-ITA
Q4_K_L	12.50 GiB	13,418,527,008	5.249	—	ManniX-ITA
Q5_K_M	14.04 GiB	15,071,090,976	5.896	—	ManniX-ITA
Q5_K_L	14.20 GiB	15,249,873,184	5.965	—	ManniX-ITA
Q6_K	16.58 GiB	17,805,297,440	6.965	—	ManniX-ITA
Q6_K_L	16.75 GiB	17,984,079,648	7.035	—	ManniX-ITA
Q8_0	19.71 GiB	21,162,527,008	8.278	—	ManniX-ITA
F16	37.06 GiB	39,789,691,168	15.565	—	ManniX-ITA

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	0.45 GiB	0.94 GiB	2.09×	5 / 25 / 0
8,192	0.61 GiB	1.88 GiB	3.10×	5 / 25 / 0
16,384	0.92 GiB	3.75 GiB	4.09×	5 / 25 / 0
32,768	1.54 GiB	7.50 GiB	4.86×	5 / 25 / 0
65,536	2.79 GiB	15.00 GiB	5.37×	5 / 25 / 0
131,072	5.29 GiB	30.00 GiB	5.67×	5 / 25 / 0

25 of 30 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 10.71 GiB. The real file is 12.33 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 7.50 GiB at 32K context where the real figure is 1.54 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

2816

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-A4B-98e-v7-coderx-it need?: Q4_K_M is exactly 13,239,744,800 bytes (12.33 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-A4B-98e-v7-coderx-it's KV cache?: 1.54 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.
Is gemma-4-A4B-98e-v7-coderx-it a mixture-of-experts model?: Yes — 98 experts, null routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of gemma-4-A4B-98e-v7-coderx-it 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.