OpenModelsHQ

H200 SXM

NVIDIA · Data center GPU

H200 SXM has 141 GB of memory and 4,800 GB/s of memory bandwidth. 420 open models in our directory fit at 4-bit.

Usable for models, before runtime overhead
141GB
Listed memory, before overhead
Total memory
141 GB
Bandwidth
4,800 GB/s
Models that fit at 4-bit
420
Within 126.9 GB, 90% of usable

NVIDIA lists the power limit as up to 700 W (configurable).

Planning a multi-GPU deployment? Check NVLink domains and network topology as well as memory fit. A GPU specification alone does not establish the interconnect in a server or cloud allocation.

What fits by category
On H200 SXM at 4-bit, 420 models fit, 3 are tight and 68 are too big, across 6 categories. Fits means the estimate is within 90% of usable memory; tight is between 90% and 100%. Models without a published size are left out.
What fits by category: data
CategoryFitsTightToo bigTotal
LLM228250280
Vision87118106
Image240024
Video220022
Speech400040
Embedding190019

Largest open models that fit at 4-bit

See all 420 in the table

A model fits when its 4-bit memory estimate (8k context) stays within 90% of usable memory, about 126.9 GB here. Showing the 30 largest of 420.

ModelTypeSizeMemory at 4-bit
Vision180B119 GB
LLM106B112.5 GB
Vision124B88.7 GB
LLM124B82.4 GB
LLM124B82.4 GB
Vision112B79.7 GB
LLM111B79.2 GB
LLM117B77.4 GB
Vision109B77.4 GB
LLM110B74.6 GB
Vision104B72.1 GB
LLM103B69.5 GB
Image83B56 GB
LLM80.4B54.2 GB
LLM78.1B51.7 GB
Vision73.4B51.4 GB
Vision72B51.3 GB
Vision71.9B51.3 GB
LLM70.6B50.3 GB
LLM70.6B50.3 GB
LLM70.6B49.5 GB
LLM70.6B49.5 GB
LLM41.8B45.3 GB
LLM67.2B45.2 GB
LLM51.6B36.8 GB
LLM51.6B36.8 GB
LLM51.6B35.2 GB
LLM51.6B35.2 GB
LLM49.1B34.7 GB
LLM40.4B28.5 GB

These are memory checks, not speed or runtime compatibility results.

Compared with

All hardware
HardwareTotal memoryUsable for modelsBandwidth
H200 SXMThis device
141 GBUp to 141 GB4,800 GB/s
80 GBUp to 80 GB2,039 GB/s
80 GBUp to 80 GB3,350 GB/s
48 GBUp to 48 GB864 GB/s

Capacity and bandwidth answer different questions: memory decides what fits, bandwidth drives how fast it runs.

Launch dates, useful-life evidence and ownership cost scenarios