llmrun

NVIDIA · GB300 Grace Blackwell Ultra Desktop Superchip · Server

Best AI Models for NVIDIA DGX Station (GB300) (252 GB)

Runs models up to 427B parameters, like MiniMax M3.

252 GB HBM7.1 TB/s2.5 PFLOPS FP161.6 kW2026

What LLMs can the NVIDIA DGX Station (GB300) run?

#ModelSpeedllmrun ScoreQuant · Memory
1
MiMo V2.6 Flash RLXiaomiHow to run →
Q4_K_M · 186.5 GB · 218 t/s
218 tok/sQ4_K_M · 186.5 GB
2
DeepSeek V4 Flash 0731DeepSeek
Q4_K_M · 183.0 GB · 169 t/s
169 tok/sQ4_K_M · 183.0 GB
3
GLM 5.3 FlashZ.ai
Q4_K_M · 205.2 GB · 205 t/s
205 tok/sQ4_K_M · 205.2 GB
4
Qwen3.8 27BAlibaba
Q4_K_M · 17.9 GB · 265 t/s
265 tok/sQ4_K_M · 17.9 GB
5
Hy3Tencent
Q4_K_M · 182.3 GB · 140 t/s
140 tok/sQ4_K_M · 182.3 GB
See all 429 models on the NVIDIA DGX Station (GB300) →Bars = traction for models without a Score yet

About the NVIDIA DGX Station (GB300)

The NVIDIA DGX Station (GB300) runs models up to 427B parameters, like MiniMax M3. A 4K-token prompt waits 110 ms before the first word on MiMo V2.6 Flash RL. GB300 Grace Blackwell Ultra Desktop Superchip (Blackwell Ultra), released 2026.

For models252 GBall of it
ChipGB300 Grace Blackwell Ultra Desktop Superchipreleased 2026

NVIDIA DGX Station (GB300) strengths and weaknesses

Strengths

  • Runs 120B MoE models: GPT OSS 120B at 366 tok/s in 70.8 GB
  • Room for long context: Llama 3.3 70B Instruct at Q4 uses 51.3 GB, leaving 200.7 GB free
  • Reads long prompts fast: A 32K-token prompt on MiMo V2.6 Flash RL starts answering in 870 ms
  • FP8 and NVFP4 run natively: Blackwell takes FP8 and NVFP4 checkpoints in vLLM as published

Memory for models

252 GB of 252 GB
35B MoE22 GB70B51.3 GB120B MoE70.8 GB427B242 GB
120B MoE70.8 GB427B242 GB
For models 252 GBModel size at Q4

Everything on the size ladder fits. MiMo V2.6 Flash RL and Qwen3.8 27B together use 204.4 GB and leave 47.6 GB free.

Similar devices and alternatives

The same chip from other makers first, then machines and GPUs people weigh against it, each with the one reason it is worth a look. Speeds on Qwen3.6 35B A3B.

Newer machine, 75% slowerMac Studio M5 Ultra (256 GB)Apple · 252.5 GB for models · M5 Ultra
▼ 75% speedsame memoryabout $9,499
Newer machine, 78% slowerMac Studio M3 Ultra (256 GB)Apple · 252.5 GB for models · M3 Ultra
▼ 78% speedsame memoryabout $7,099
Newer machine, 79% slowerMac Studio M2 Ultra (192 GB)Apple · 188.5 GB for models · M2 Ultra
▼ 78% speed▼ 63.5 GBabout $5,599
Step down to 124.5 GBMac Studio M4 Max (128 GB)Apple · 124.5 GB for models · M4 Max
▼ 82% speed▼ 127.5 GBabout $3,499
Step down to 24 GBRTX 4090NVIDIA · 24 GB VRAM · GPU
▼ 57% speed▼ 228 GBabout $1,599rent $0.34/hr
Step down to 32 GBRTX 5090NVIDIA · 32 GB VRAM · GPU
▼ 40% speed▼ 220 GBabout $5,502rent $0.50/hr

NVIDIA DGX Station (GB300) tokens per second by model size

Output speed depends on memory bandwidth (7100 GB/s). Grey rows need more than the 252 GB available.

  1. 9BQwen3.5 9B · Q4_K_M · 6.6 GB
    726tok/s
  2. 14BQwen3 14B · Q4_K_M · 11.9 GB
    486tok/s
  3. 27–31BQwen3.8 27B · Q4_K_M · 17.9 GB
    265tok/s
  4. 35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB
    354tok/s
  5. 70BLlama 3.3 70B Instruct · Q4_K_M · 51.3 GB
    99.1tok/s
  6. 120B MoEGPT OSS 120B · Q4_K_M · 70.8 GB
    366tok/s

Estimated speeds: Q4_K_M, 4K context, one user. How we estimate

Unified-memory platforms: writing vs compute

Writing, tok/sCompute, TFLOPS
  1. M2 Ultra192 GB · 800 GB/s
    75.9
    54
  2. M3 Ultra256 GB · 819 GB/s
    76.6
    66
  3. M5 Ultra256 GB · 1228 GB/s
    87.9
    109
  4. GB300 Grace Blackwell Ultra Desktop Superchip252 GB · 7100 GB/s
    354
    2500

Machines with about the same memory bandwidth write at about the same speed; compute decides how fast prompts are read. Speeds on Qwen3.6 35B A3B.

All 1346 models on the NVIDIA DGX Station (GB300)

Featured only

Best quant that fits in 252 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works

LLM models compatible with NVIDIA DGX Station (GB300) — ranked by Score, recency and Fit
ModelVRAMScoreFit
Olmo 3.1 32B Think
BF16·70.7 t/s
66.6 GBS86, Runs great
Qwopus3.5 9B V3.5
Q4_K_M·725.6 t/s
6.6 GBS100, Runs great
Hermes 4 14B
Q4_K_M·485.8 t/s
11.8 GBS100, Runs great
GLM 5
IQ2_XXS·145.6 t/s
209.1 GBS93, Runs great
Ouro 2.6B Thinking
BF16·716.6 t/s
12.1 GBS100, Runs great
Hy MT2 1.8B
Q4_K_M·2780.1 t/s
2.6 GBS100, Runs great
Nafie 473M
BF16·4437.5 t/s
1.0 GBS100, Runs great
CHADROCK3.6 35B UNCENSORED MTP STRIX LEAN
BF16·59.9 t/s
80.9 GBA84, Runs well
Functiongemma 270M IT
BF16·7822.0 t/s
1.6 GBS100, Runs great
StorySupra 10M
Q4_K_M·14421.9 t/s
0.3 GBS100, Runs great
Tiny Aya Water
BF16·626.2 t/s
9.4 GBS100, Runs great
Qwen3 Next Moe
Q4_K_M·15383.3 t/s
0.3 GBS100, Runs great
PicoMistral 23M
Q4_K_M·14421.9 t/s
0.3 GBS100, Runs great
Ouro Hybrid 1.4B
BF16·1237.3 t/s
6.5 GBS100, Runs great
Supra Mini V6 1M
Q4_K_M·15383.3 t/s
0.3 GBS100, Runs great

NVIDIA DGX Station (GB300) specs and power

Specs

Chip
GB300 Grace Blackwell Ultra Desktop Superchip · NVIDIA Blackwell Ultra (GB300 Desktop)
CPU cores
72
Memory
252 GB
For models
252 GB
Memory bandwidth
7.1 TB/s
FP16 (dense)
2.5 PFLOPS
Power
1.6 kW max
Networking
ConnectX-8 SuperNIC up to 800 Gb/s
OS
DGX OS (Ubuntu, Arm64); Windows version announced, no date
Released
2026-08-01

NVIDIA DGX Station (GB300) FAQ

How much memory is available for AI on the NVIDIA DGX Station (GB300)?

All 252 GB can hold model weights.

What is the largest model the NVIDIA DGX Station (GB300) can run?

MiniMax M3 (427.0B) is the largest at 4-bit or better: at Q4_K_S it uses 242.0 GB, 96% of the available memory, and runs at about 175 tok/s.

Can the NVIDIA DGX Station (GB300) run MiMo V2.6 Flash RL?

Yes. MiMo V2.6 Flash RL fits at Q4_K_M, using 186.5 GB of the 252 GB available, and it is the best model that runs well on this machine.

How long does the NVIDIA DGX Station (GB300) take to read a long prompt?

On MiMo V2.6 Flash RL, about 110 ms for a 4K-token prompt and 870 ms for a 32K-token document before the first word appears.

Related

48 GB GPUsServersBest coding LLMsAll hardware