How much VRAM does AMALIA-9B-0626-DPO need?

Q4_K_M is exactly 5,582,838,528 bytes (5.20 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is AMALIA-9B-0626-DPO's KV cache?

5.25 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

amalia-llm · text

AMALIA-9B-0626-DPO

amalia-llm/AMALIA-9B-0626-DPO

AMALIA-9B-0626-DPO at Q4_K_M is exactly 5,582,838,528 bytes (5.20 GiB / 5.58 GB) — an effective 4.880 bits per weight, not the nominal 4. Its KV cache at 32K is 5.25 GiB.

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

Parameters

9.2B

Architecture

llama

42 layers

Context

32,768

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	3.35 GiB	3,593,067,264	3.141	—	duarteocarmo
Q3_K_S	3.86 GiB	4,141,767,424	3.620	—	duarteocarmo
Q3_K_M	4.24 GiB	4,553,136,896	3.980	—	duarteocarmo
Q3_K_L	4.58 GiB	4,913,847,040	4.295	—	duarteocarmo
IQ4_XS	4.74 GiB	5,084,863,232	4.445	—	duarteocarmo
Q4_K_S	4.96 GiB	5,321,186,048	4.651	—	duarteocarmo
Q4_K_S	4.96 GiB	5,321,186,176	4.651	—	layerx-labs
Q4_K_M	5.20 GiB	5,582,838,528	4.880	—	duarteocarmo
Q4_K_M	5.20 GiB	5,582,838,592	4.880	—	csoares31
Q4_K_M	5.20 GiB	5,582,838,624	4.880	—	layerx-labs
Q5_K_S	5.93 GiB	6,366,092,032	5.565	—	duarteocarmo
Q5_K_M	6.07 GiB	6,518,168,320	5.697	—	duarteocarmo
Q6_K	7.00 GiB	7,511,956,224	6.566	—	duarteocarmo
Q8_0	9.06 GiB	9,728,449,280	8.504	—	duarteocarmo
Q8_0	9.06 GiB	9,728,449,344	8.504	—	csoares31
Q8_0	9.06 GiB	9,728,449,376	8.504	—	layerx-labs
BF16	17.05 GiB	18,308,422,400	16.003	—	duarteocarmo
F16	17.05 GiB	18,308,422,464	16.003	—	csoares31
BF16	17.05 GiB	18,308,422,528	16.003	—	layerx-labs

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.66 GiB	0.66 GiB	—	42 / 0 / 0
8,192	1.31 GiB	1.31 GiB	—	42 / 0 / 0
16,384	2.63 GiB	2.63 GiB	—	42 / 0 / 0
32,768	5.25 GiB	5.25 GiB	—	42 / 0 / 0
65,536	10.50 GiB	10.50 GiB	—	42 / 0 / 0
131,072	21.00 GiB	21.00 GiB	—	42 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs gemma-4-12B-it-qat-q4_0-unquantized vs gemma-4-E4B-it vs Qwen3-30B-A3B-Thinking-2507 vs Qwen3-4B vs Qwen3-8B 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 4.79 GiB. The real file is 5.20 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

4096

Vocab

128,000

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does AMALIA-9B-0626-DPO need?: Q4_K_M is exactly 5,582,838,528 bytes (5.20 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is AMALIA-9B-0626-DPO's KV cache?: 5.25 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 AMALIA-9B-0626-DPO 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.