GPU GUIDE · NVIDIA

Best AI models for the
NVIDIA DGX Station (Blackwell Ultra)

The NVIDIA DGX Station (Blackwell Ultra) has 784GB of unified memory. Below are the top 30 open-source AI models that fit, ranked by composite benchmark score. Each row shows the best quantization that fits your hardware.

VRAM
784GB
Brand
nvidia
Models that fit
30
Generation
DGX

Top 30 models for the NVIDIA DGX Station (Blackwell Ultra)

01

MiMo V2.5 Pro

320B
89.7
Best fit: fp16 · 641.0GBCan I run it? →
02

Qwen 3.6 Max

480B
86.4
Best fit: Q8_0 · 511.0GBCan I run it? →
03

DeepSeek V4 Pro

685B
85.8
Best fit: Q8_0 · 728.8GBCan I run it? →
04

Qwen 3.6 Plus

235B
83.3
Best fit: fp16 · 471.0GBCan I run it? →
05

GLM-5

230B
82.9
Best fit: fp16 · 461.0GBCan I run it? →
06

MiniMax M2.7

456B
82.7
Best fit: Q8_0 · 485.5GBCan I run it? →
07

DeepSeek V4 Flash

37.0B
77.5
Best fit: fp16 · 75.0GBCan I run it? →
08

Qwen 3.6 27B

27.0B
76.4
Best fit: fp16 · 55.0GBCan I run it? →
09

Qwen 3.5 397B A17B

397B
75.1
Best fit: Q8_0 · 422.8GBCan I run it? →
10

MiMo V2 Omni

120B
74.9
Best fit: fp16 · 241.0GBCan I run it? →
11

Qwen 3.6 35B A3B

35.0B
72.5
Best fit: fp16 · 71.0GBCan I run it? →
12

Qwen3.5-27B

27.8B
70.1
Best fit: f32 · 112.2GBCan I run it? →
13

DeepSeek V3.2

685B
69.5
Best fit: Q8_0 · 728.8GBCan I run it? →
14

Qwen 3.5 122B A10B

122B
69.3
Best fit: fp16 · 245.0GBCan I run it? →
15

Mistral Medium 3.5 128B

128B
65.4
Best fit: f32 · 511.8GBCan I run it? →
16

Gemma 4 31B (free)

31.0B
65.3
Best fit: fp16 · 63.0GBCan I run it? →
17

Qwen 3.5 Omni Plus

235B
64.4
Best fit: fp16 · 471.0GBCan I run it? →
18

Qwen 3.5 35B A3B

35.0B
61.9
Best fit: fp16 · 71.0GBCan I run it? →
19

NVIDIA Nemotron 3 Super

340B
60.0
Best fit: fp16 · 681.0GBCan I run it? →
20

NVIDIA Nemotron 3 Super 120B A12B BF16

124B
59.9
Best fit: fp16 · 248.2GBCan I run it? →
21

GPT-OSS 120B

120B
55.4
Best fit: fp16 · 241.0GBCan I run it? →
22

Qwen 3.5 9B

9.0B
54.0
Best fit: fp16 · 19.0GBCan I run it? →
23

gemma 4 31B

32.7B
53.8
Best fit: fp16 · 66.4GBCan I run it? →
24

Gemma 4 26B A4B (free)

26.5B
52.0
Best fit: f32 · 107.0GBCan I run it? →
25

Qwen3 235B A22B Thinking 2507

235B
49.2
Best fit: Q6_K · 194.7GBCan I run it? →
26

Qwen 3 Coder

480B
46.7
Best fit: Q8_0 · 511.0GBCan I run it? →
27

Mistral Small 4

119B
46.3
Best fit: f32 · 478.6GBCan I run it? →
28

Qwen3 VL 235B A22B Instruct

236B
46.1
Best fit: fp16 · 472.4GBCan I run it? →
29

Qwen 3.5 4B

4.0B
45.1
Best fit: fp16 · 9.0GBCan I run it? →
30

DeepSeek R1 0528

685B
45.1
Best fit: Q8_0 · 728.8GBCan I run it? →
Advertisement

FAQ — running AI on the NVIDIA DGX Station (Blackwell Ultra)

How many AI models can the NVIDIA DGX Station (Blackwell Ultra) run?

With 784GB of unified memory, the NVIDIA DGX Station (Blackwell Ultra) can run 30+ open-source models from our database, including MiMo V2.5 Pro, Qwen 3.6 Max, DeepSeek V4 Pro.

What's the largest LLM I can run on a NVIDIA DGX Station (Blackwell Ultra)?

The biggest model that fits is approximately 685B. Larger models would need to be quantized further or won't fit at all.

Is 784GB of unified memory enough for local AI?

Yes — 784GB comfortably runs most popular open-source models including 30B-class LLMs at Q4_K_M.