How much VRAM does Hermes-4.3-36B need?

Q4_K_M is exactly 21,762,145,888 bytes (20.27 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Hermes-4.3-36B's KV cache?

8.00 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.

NousResearch · text

Hermes-4.3-36B

NousResearch/Hermes-4.3-36B

Hermes-4.3-36B at Q4_K_M is exactly 21,762,145,888 bytes (20.27 GiB / 21.76 GB) — an effective 4.816 bits per weight, not the nominal 4. Its KV cache at 32K is 8.00 GiB.

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

Parameters

36.2B

Architecture

seed_oss

64 layers

Context

524,288

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ2_XXS	9.23 GiB	9,910,923,104	2.193	—	bartowski
IQ2_XS	10.19 GiB	10,945,081,184	2.422	—	bartowski
IQ2_S	10.82 GiB	11,612,626,784	2.570	—	bartowski
IQ2_M	11.68 GiB	12,541,927,264	2.775	—	bartowski
Q2_K	12.67 GiB	13,604,183,648	3.010	—	MaziyarPanahi
Q2_K	12.67 GiB	13,604,183,904	3.010	—	bartowski
IQ3_XXS	13.15 GiB	14,116,757,344	3.124	—	bartowski
Q2_K_L	13.39 GiB	14,379,863,904	3.182	—	bartowski
IQ3_XS	14.05 GiB	15,089,946,464	3.339	—	bartowski
Q3_K_S	14.77 GiB	15,855,406,944	3.509	—	bartowski
IQ3_M	15.36 GiB	16,496,021,344	3.651	—	bartowski
Q3_K_M	16.41 GiB	17,620,946,528	3.899	—	MaziyarPanahi
Q3_K_M	16.41 GiB	17,620,946,784	3.899	—	bartowski
Q3_K_L	17.83 GiB	19,142,692,448	4.236	—	MaziyarPanahi
Q3_K_L	17.83 GiB	19,142,692,704	4.236	—	bartowski
IQ4_XS	18.16 GiB	19,498,614,624	4.315	—	bartowski
IQ4_NL	19.18 GiB	20,592,983,904	4.557	—	bartowski
Q4_0	19.21 GiB	20,621,819,744	4.564	—	bartowski
Q4_K_S	19.27 GiB	20,695,220,064	4.580	—	bartowski
Q4_K_M	20.27 GiB	21,762,145,888	4.816	—	MaziyarPanahi
Q4_K_M	20.27 GiB	21,762,146,144	4.816	—	bartowski
Q4_K_L	20.82 GiB	22,351,662,944	4.946	—	bartowski
Q4_1	21.20 GiB	22,760,750,944	5.037	—	bartowski
Q5_K_S	23.26 GiB	24,970,461,024	5.526	—	bartowski
Q5_K_M	23.84 GiB	25,594,363,488	5.664	—	MaziyarPanahi
Q5_K_M	23.84 GiB	25,594,363,744	5.664	—	bartowski
Q5_K_L	24.29 GiB	26,084,593,504	5.772	—	bartowski
Q6_K	27.63 GiB	29,666,094,688	6.565	—	MaziyarPanahi
Q6_K	27.63 GiB	29,666,094,944	6.565	—	bartowski
Q6_K_L	27.99 GiB	30,050,832,224	6.650	—	bartowski
Q8_0	35.78 GiB	38,421,090,144	8.502	—	bartowski
BF162 shards	67.35 GiB	72,311,394,080	16.002	—	bartowski

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	1.00 GiB	1.00 GiB	—	64 / 0 / 0
8,192	2.00 GiB	2.00 GiB	—	64 / 0 / 0
16,384	4.00 GiB	4.00 GiB	—	64 / 0 / 0
32,768	8.00 GiB	8.00 GiB	—	64 / 0 / 0
65,536	16.00 GiB	16.00 GiB	—	64 / 0 / 0
131,072	32.00 GiB	32.00 GiB	—	64 / 0 / 0

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 18.94 GiB. The real file is 20.27 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

5120

Vocab

155,136

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Hermes-4.3-36B need?: Q4_K_M is exactly 21,762,145,888 bytes (20.27 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Hermes-4.3-36B's KV cache?: 8.00 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 Hermes-4.3-36B 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.