NVIDIA DGX Station for Windows (GB300)
NVIDIAannounced
Announced, not on sale yet: every number here is the maker's promise.
On this page
What it can run
The agent count is a memory estimate for this model and context. The speed check below is a separate estimate, not a measured rate per agent.
How it is calculated
How many agents of each open model fit in its memory at once, each with room for about 32 thousand tokens of conversation — a typical coding agent. The agents share one copy of the model, so the first one needs the most memory and each next adds only its own conversation. Our estimate from memory size; open a row to try other settings.
- GPT-OSS-120B4-bit, as releasedFirst agent: 68.1 GBeach next +6.5 GBAgents at once: 29
Details
each next +6.5 GB
- GPT-OSS-20B4-bit, as releasedFirst agent: 16.6 GBeach next +3.8 GBAgents at once: 63
Details
each next +3.8 GB
- Qwen3-Coder-30B-A3B4-bit, smallerFirst agent: 22.3 GBeach next +3.9 GBAgents at once: 59
Details
each next +3.9 GB
- Qwen3-Coder-30B-A3B8-bit, closer to the originalFirst agent: 36.0 GBeach next +4.5 GBAgents at once: 49
Details
each next +4.5 GB
Available to models: 252 GB of 748 GB. The maker lets the graphics chip use this much.
Measured speed
Specification
- Memory
- Of it, for the graphics chip
- Memory speed
- claimed
- Power draw
- not in our data yet
- Combines with others
- not in our data yet
- Price
- no confirmed price in our data yet
How the fit score is made
Announced machines get no score until they go on sale and can be measured.