How much VRAM does UI-TARS-72B-DPO need?

Q4_K_M is exactly 47,415,712,640 bytes (44.16 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is UI-TARS-72B-DPO'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.

ByteDance-Seed · vision language

UI-TARS-72B-DPO

ByteDance-Seed/UI-TARS-72B-DPO

UI-TARS-72B-DPO at Q4_K_M is exactly 47,415,712,640 bytes (44.16 GiB / 47.42 GB) — an effective 5.168 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

73.4B

Architecture

qwen2vl

80 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
IQ1_M	22.11 GiB	23,740,210,304	2.587	—	bartowski
IQ2_XXS	23.74 GiB	25,490,021,504	2.778	—	bartowski
IQ2_XS	25.20 GiB	27,057,642,624	2.949	—	bartowski
IQ2_S	26.02 GiB	27,939,134,592	3.045	—	bartowski
IQ2_M	27.32 GiB	29,338,983,552	3.197	—	bartowski
Q2_K	27.76 GiB	29,811,760,256	3.249	—	bartowski
Q2_K_L	28.90 GiB	31,028,272,256	3.382	—	bartowski
IQ3_XXS	29.66 GiB	31,845,080,192	3.471	—	bartowski
Q3_K_S	32.12 GiB	34,487,786,624	3.759	—	bartowski
IQ3_M	33.07 GiB	35,503,594,624	3.869	—	bartowski
Q3_K_M	35.11 GiB	37,698,722,944	4.109	—	bartowski
Q3_K_L	36.79 GiB	39,505,222,528	4.305	—	lmstudio-community
Q3_K_L	36.79 GiB	39,505,222,784	4.305	—	bartowski
IQ4_XS	36.98 GiB	39,709,072,512	4.328	—	bartowski
IQ4_NL	38.48 GiB	41,315,622,016	4.503	—	bartowski
Q4_0	38.54 GiB	41,383,124,096	4.510	—	bartowski
Q4_K_S	40.88 GiB	43,889,220,736	4.783	—	bartowski
Q4_1	42.56 GiB	45,697,883,264	4.980	—	bartowski
Q4_K_M	44.16 GiB	47,415,712,640	5.168	—	lmstudio-community
Q4_K_M	44.16 GiB	47,415,712,896	5.168	—	bartowski
Q5_K_S2 shards	47.85 GiB	51,375,136,064	5.599	—	bartowski
Q5_K_M2 shards	50.71 GiB	54,447,463,776	5.934	—	bartowski
Q6_K2 shards	59.93 GiB	64,347,626,592	7.013	—	lmstudio-community
Q6_K2 shards	59.93 GiB	64,347,626,848	7.013	—	bartowski
Q8_02 shards	71.96 GiB	77,262,609,504	8.420	—	lmstudio-community
Q8_02 shards	71.96 GiB	77,262,609,760	8.420	—	bartowski

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 Qwen3.6-35B-A3B vs Qwen3.5-9B vs gemma-4-26B-A4B-it vs Qwythos-9B-Claude-Mythos-5-1M vs Qwen3-VL-30B-A3B-Instruct vs gemma-4-26B-A4B-it-qat-q4_0-unquantized 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 38.45 GiB. The real file is 44.16 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

152,064

Sliding window

32768

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

false

This model declares a sliding window but sets use_sliding_window: false, so the window is not applied. Honouring the field without the flag understates KV for the whole family.

Questions people ask

How much VRAM does UI-TARS-72B-DPO need?: Q4_K_M is exactly 47,415,712,640 bytes (44.16 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is UI-TARS-72B-DPO'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 UI-TARS-72B-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.