The models we advise on
We advise on three open-weight models. All three are MIT licensed, which means there's no usage restriction or per-token fee from the model creator — the only cost is the hardware to run them. Below is the exact memory math for each one, and the smallest build in our catalog that can actually hold it.
DeepSeek V4 Flash
284 billion parameters · MIT license
Minimum build that meets this
4x RTX PRO 6000 Blackwell = 384 GB combined GPU memory — above the 341 GB requirement. Three cards would only reach 288 GB, short of what's needed to even load the model, so four is the minimum. This is our Small Startup build — $53,000, 2,400 W.
GLM 5.2
744 billion parameters · MIT license
Minimum build that meets this
A single H200 SXM (141 GB) or even several of them wired ad hoc isn't a product we sell as a coherent system — the smallest complete system in our catalog that clears 893 GB is one DGX B200: 1,440 GB combined, already NVLink-connected out of the box, well above the requirement. This is our Mid-Size Company build — $515,000, 14,300 W.
DeepSeek V4 Pro
1,600 billion parameters · MIT license
Minimum build that meets this
A single DGX B200 (1,440 GB) is not enough on its own — it falls short of the 1,920 GB requirement. Two DGX B200 systems combined reach 2,880 GB, which clears it. See the full breakdown, including the GB300 NVL72 rack alternative — $1,030,000, 28,600 W for the 2x DGX B200 option.
Ready to talk hardware? Head to Help Me Choose for a guided recommendation, or go straight to Request a Quote.