Sunk Cost sunkcost.ai Data checked 2026-09-03

Can a NVIDIA GeForce RTX 5090, 32GB run local LLMs?

Yes — 27 of the 39 open models on this site fit in its 31 GB of usable memory, the strongest being Qwen3.8 27B. Whether that saves you money is a different question, and the answer is usually no.

Price$1,999
Memory32 GB, about 31 GB of it addressable by the GPU at 1792 GB/s
Best model it runsQwen3.8 27B — Sonnet-class, 66 tok/s
Pay-back against the APIPays back in 25 years at 500k tokens a day

Run the numbers on this machine

What it runs

ModelSpeedClassGood atMemory
Qwen3.8 27BQ4_K_M 66 tok/s Sonnet-class 19 GB
Qwen3.6 27BQ4_K_M 65 tok/s Haiku-class 19 GB
Qwen3.6 35B-A3BQ4_K_M 235 tok/s Haiku-class 23 GB
Muse Glimmer 30BQ4_K_M 69 tok/s Haiku-class 18 GB
Gemma 4 26B-A4BQ4_K_M 172 tok/s Haiku-class 18 GB
Gemma 4 31B itQ4_K_M 48 tok/s Haiku-class 26 GB
Granite 4.2 30BQ4_K_M 47 tok/s Haiku-class 26 GB
GLM-4.7-FlashQ4_K_M 167 tok/s Haiku-class 20 GB
Gemma 4 12BQ4_K_M 155 tok/s Below every hosted tier 8.0 GB
Qwen3.5 9BQ4_K_M 183 tok/s Below every hosted tier 6.8 GB
Nemotron 3.5 Lightning 30B-A3BQ4_K_M 226 tok/s Below every hosted tier 26 GB
MiniCPM5 2BQ4_K_M 416 tok/s Below every hosted tier 3.0 GB

15 more fit; the calculator lists them all.

The specifics

Chip
GeForce RTX 5090 — Blackwell · 32 GB GDDR7
Memory bandwidth
1792 GB/s
Usable by the GPU
31 GB — Bandwidth: 28 Gbps × 512-bit bus ÷ 8 = 1,792 GB/s, which NVIDIA's compare page also publishes. Usable memory is the VRAM less 1 GB: the margin llama.cpp's automatic fitting (--fit) leaves free on each device by default (common/common.h, fit_params_target = 1 GiB). Treated as nominal GB, like the other machines here.
Power under load
575 W (published) — 575 W is the maker's rated board power for the card alone. The rest of the PC adds more, and LLM decoding usually draws less than the full rating. NVIDIA recommends a 1000 W system power supply.
Availability
Launch price. Founders Edition out of stock at NVIDIA when checked 2026-09-16; partner cards sell for more.
Sources
source 1, source 2, source 3, source 4, source 5, source 6