How much VRAM does Apertus-70B-Instruct-2509 need?

Q4_K_M is exactly 43,721,584,608 bytes (40.72 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Apertus-70B-Instruct-2509's KV cache?

10.00 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.

swiss-ai · text

Apertus-70B-Instruct-2509

swiss-ai/Apertus-70B-Instruct-2509

Apertus-70B-Instruct-2509 at Q4_K_M is exactly 43,721,584,608 bytes (40.72 GiB / 43.72 GB) — an effective 4.954 bits per weight, not the nominal 4. Its KV cache at 32K is 10.00 GiB.

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

Parameters

70.6B

Architecture

apertus

80 layers

Context

65,536

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
UD-IQ1_S	15.46 GiB	16,599,131,808	1.881	—	unsloth
IQ1_M	15.74 GiB	16,903,218,976	1.915	—	bartowski
UD-IQ1_M	16.58 GiB	17,799,227,040	2.017	—	unsloth
IQ2_XXS	17.88 GiB	19,203,532,576	2.176	—	bartowski
UD-IQ2_XXS	18.54 GiB	19,908,437,664	2.256	—	unsloth
IQ2_XS	19.75 GiB	21,211,555,616	2.404	—	bartowski
IQ2_S	20.89 GiB	22,433,408,800	2.542	—	bartowski
IQ2_M	22.61 GiB	24,273,659,680	2.751	—	bartowski
UD-IQ2_M	23.12 GiB	24,825,341,600	2.813	—	unsloth
Q2_K	25.40 GiB	27,272,048,608	3.090	—	giladgd
Q2_K	25.40 GiB	27,272,062,624	3.090	—	unsloth
Q2_K	25.40 GiB	27,272,062,752	3.090	—	bartowski
IQ3_XXS	25.53 GiB	27,411,523,360	3.106	—	bartowski
Q2_K_L	25.63 GiB	27,523,720,864	3.119	—	unsloth
UD-IQ3_XXS	26.01 GiB	27,932,665,504	3.165	—	unsloth
Q2_K_L	26.38 GiB	28,320,638,752	3.209	—	bartowski
IQ3_XS	27.55 GiB	29,583,124,256	3.352	—	bartowski
Q3_K_S	28.65 GiB	30,768,000,992	3.486	—	giladgd
Q3_K_S	28.65 GiB	30,768,015,008	3.486	—	unsloth
Q3_K_S	28.65 GiB	30,768,015,136	3.486	—	bartowski
IQ3_M	29.84 GiB	32,038,102,816	3.630	—	bartowski
Q3_K_M	33.10 GiB	35,535,876,064	4.027	—	giladgd
Q3_K_M	33.10 GiB	35,535,890,080	4.027	—	unsloth
Q3_K_M	33.10 GiB	35,535,890,208	4.027	—	bartowski
IQ4_XS	35.33 GiB	37,933,983,520	4.298	—	bartowski
IQ4_XS	35.36 GiB	37,967,537,824	4.302	—	unsloth
Q3_K_L	36.87 GiB	39,591,768,032	4.486	—	giladgd
Q3_K_L	36.87 GiB	39,591,782,176	4.486	—	bartowski
Q4_0	37.25 GiB	40,001,761,248	4.533	—	giladgd
IQ4_NL	37.33 GiB	40,085,661,344	4.542	—	unsloth
IQ4_NL	37.33 GiB	40,085,661,472	4.542	—	bartowski
Q4_0	37.46 GiB	40,221,976,224	4.558	—	unsloth
Q4_0	37.46 GiB	40,221,976,352	4.558	—	bartowski
Q4_K_S	37.67 GiB	40,446,357,472	4.583	—	giladgd
Q4_K_S	37.67 GiB	40,446,371,488	4.583	—	unsloth
Q4_K_S	37.67 GiB	40,446,371,616	4.583	—	bartowski
Q4_K_M	40.72 GiB	43,721,584,608	4.954	—	giladgd
Q4_K_M	40.72 GiB	43,721,598,624	4.954	—	unsloth
Q4_K_M	40.72 GiB	43,721,598,752	4.954	—	bartowski
Q4_1	41.30 GiB	44,347,074,208	5.025	—	unsloth

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	1.25 GiB	1.25 GiB	—	80 / 0 / 0
8,192	2.50 GiB	2.50 GiB	—	80 / 0 / 0
16,384	5.00 GiB	5.00 GiB	—	80 / 0 / 0
32,768	10.00 GiB	10.00 GiB	—	80 / 0 / 0
65,536	20.00 GiB	20.00 GiB	—	80 / 0 / 0
131,072	40.00 GiB	40.00 GiB	—	80 / 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 36.99 GiB. The real file is 40.72 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

8192

Vocab

131,072

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Apertus-70B-Instruct-2509 need?: Q4_K_M is exactly 43,721,584,608 bytes (40.72 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Apertus-70B-Instruct-2509's KV cache?: 10.00 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 Apertus-70B-Instruct-2509 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.