Local AI Deals

Cheapest box to run DeepSeek-V4-Flash at Q6_K

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 finderprices checked 33 days ago
Quant
Context

DeepSeek-V4-Flash at Q6_K needs 234GB.

Cheapest Config that fits: Mac Studio, M5 Ultra, 512GB. Fits, 270GB headroom. Est. 56.3 tok/s.

No price yet: Ships late October 2026. Apple has not published a price.

Price history: 34 days recorded.

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.

Offers — Mac Studio, M5 Ultra, 512GB, 1TB

Ships late October 2026. Apple has not published a price.

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 × 0.820 B/param (6.56 bits) + 0.006 GB/1K × 8K K/V + 1.2 GB runtime
speed = bandwidth × 0.5 ÷ (13B active × 0.820 B/param) = 10.7 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.

Email me when Mac Studio, M5 Ultra, 512GB is back in stock. One email, then we stop.

No account. One-click unsubscribe in the email.

What else Mac Studio M5 Ultra 512GB runs

Best quant that still fits at 8K context, largest model first. Full list on the Config page.

Bandwidth 1.2TB/s, usable memory 504GB of 512GB.