Local AI Deals

Cheapest box to run DeepSeek-V4-Flash at BF16

DeepSeek’s DeepSeek-V4-Flash is 284B parameters, 13B of them read per token. The number below is computed from that and each vendor’s published specs, not asserted. Switch the quant or the context and it recomputes in place.

Fit finderspecs, not prices
Quant
Context

DeepSeek-V4-Flash at BF16 needs 569GB.

Nothing on the Grid runs DeepSeek-V4-Flash at BF16 today. The largest Config, Mac Studio, M5 Ultra, 512GB, is 65GB short. At it fits with 108GB headroom.

Compressed sparse attention. Each layer stores one latent, divided by the compression ratio in its config (mostly 4x or 128x); the figure matches DeepSeek's own claim of about a tenth of the V3 cache.

GeForce RTX 5090 is a card, not a computer: every price shown for it is the cheapest listing we found plus $1,000 for the minimum host machine.

memory = 284B × 2.000 B/param (16 bits) + 0.006 GB/1K × 8K K/V + 1.2 GB runtime
speed = bandwidth × 0.5 ÷ (13B active × 2.000 B/param) = 26.0 GB read per token
usable = memory − operating system reserve (8% on a Mac, 6% on a DGX Spark, 4% on a card)

Bandwidth is the vendor’s published figure. Every speed is an estimate and says so. Methodology.