DeepSeek V4 Flash

MIT License

Smallest footprint, lowest build cost. Best starting point if you're not sure how much scale you need.

MoE · Total params: 284B · Active params: 13B

Needs 341 GB GPU memory to run

284 × 1 GB + 20% = 341 GB

Verified builds

3× H200 NVL

423GB

✓ covers DeepSeek V4 Flash's 341 GB

$107,994

VERIFIED

1.8 kW→1.5homes

Request a quote

1× DGX H200

1,128GB

✓ covers DeepSeek V4 Flash's 341 GB

~$400–500K

VERIFIED EXAMPLE

$467,311 (Broadberry) — Source: Broadberry

10.2 kW→8.5homes·2.7EV batteries/day

Request a quote

What about 11× RTX 5090 cards at $45,089?

Desktop cards have no NVLink. 11 cards = eleven separate 32 GB islands, not one 352 GB pool. The model needs a single contiguous 341 GB space. This pile cannot run it.

The cluster that can: 3× H200 NVL at $107,994.