FR
← All hardware
Datacenter · HBM3e

B200 192 GB: which LLMs can it run?

On its own, it loads up to Qwen3 235B-A22B in Q4. With several cards, memory adds up and opens bigger models.

Specifications

Memory192 GB
Bandwidth8,000 GB/s
FP16 compute~2,250 TF
Power1000 W
Indicative price€42,000
Cloud rental€5.50/h
Build a server with it →Compare

Against cards at the same price

B200 192 GB vs H200 141 GB€33,000
B200 192 GB vs H100 80 GB€28,000
B200 192 GB vs Mac Studio M3 Ultra 512 GB€11,500

Models and expected speed

ModelQ4Q8FP16
Qwen3 4B✓ ~1,725 tok/s✓ ~1,002 tok/s✓ ~540 tok/s
Llama 3.1 8B✓ ~889 tok/s✓ ~510 tok/s✓ ~272 tok/s
Qwen3 8B✓ ~868 tok/s✓ ~498 tok/s✓ ~266 tok/s
Gemma 3 12B✓ ~599 tok/s✓ ~342 tok/s✓ ~182 tok/s
Mistral Nemo 12B✓ ~589 tok/s✓ ~336 tok/s✓ ~179 tok/s
Phi-4 14B✓ ~491 tok/s✓ ~280 tok/s✓ ~149 tok/s
Qwen3 14B✓ ~487 tok/s✓ ~278 tok/s✓ ~148 tok/s
gpt-oss-20b✓ ~1,905 tok/s✓ ~1,109 tok/s✓ ~599 tok/s
Mistral Small 24B✓ ~302 tok/s✓ ~172 tok/s✓ ~91 tok/s
Gemma 3 27B✓ ~269 tok/s✓ ~153 tok/s✓ ~81 tok/s
Qwen3.8 27B✓ ~260 tok/s✓ ~148 tok/s✓ ~78 tok/s
Qwen3 30B-A3B✓ ~2,066 tok/s✓ ~1,206 tok/s✓ ~652 tok/s
Qwen3 32B✓ ~222 tok/s✓ ~126 tok/s✓ ~67 tok/s
Qwen2.5 Coder 32B✓ ~222 tok/s✓ ~126 tok/s✓ ~67 tok/s
DeepSeek R1 Distill 32B✓ ~222 tok/s✓ ~126 tok/s✓ ~67 tok/s
Mixtral 8x7B✓ ~558 tok/s✓ ~318 tok/s✓ ~170 tok/s
Llama 3.3 70B✓ ~104 tok/s✓ ~59 tok/s✓ ~31 tok/s
GLM-4.5-Air 106B✓ ~599 tok/s✓ ~342 tok/s2 × · ~273 tok/s
gpt-oss-120b✓ ~1,371 tok/s✓ ~792 tok/s2 × · ~638 tok/s
Mistral Large 123B✓ ~59 tok/s✓ ~34 tok/s2 × · ~27 tok/s
Qwen3 235B-A22B✓ ~330 tok/s2 × · ~281 tok/s3 × · ~224 tok/s
GLM-4.5 355B2 × · ~341 tok/s3 × · ~291 tok/s5 × · ~257 tok/s
Llama 3.1 405B2 × · ~27 tok/s3 × · ~23 tok/s5 × · ~20 tok/s
DeepSeek R1 671B3 × · ~443 tok/s5 × · ~419 tok/s8 × · ~356 tok/s
DeepSeek V3.1 671B3 × · ~443 tok/s5 × · ~419 tok/s8 × · ~356 tok/s
Kimi K2 1T4 × · ~682 tok/s6 × · ~581 tok/s✗

8k tokens, one user. “2 ×” = number of cards needed, speed with tensor parallelism (vLLM). ✗ = doesn’t fit, even with 8 cards.