Local AI Deals

Cheapest box to run Llama-4-Maverick-17B-128E at BF16

Meta’s Llama-4-Maverick-17B-128E is 400B parameters, 17B 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

Llama-4-Maverick-17B-128E at BF16 needs 803GB.

Nothing on the Grid runs Llama-4-Maverick-17B-128E at BF16 today. The largest Config, Mac Studio, M5 Ultra, 512GB, is 299GB short. At it fits with 50GB headroom.

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 = 400B × 2.000 B/param (16 bits) + 0.197 GB/1K × 8K K/V + 1.2 GB runtime
speed = bandwidth × 0.5 ÷ (17B active × 2.000 B/param) = 34.0 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.