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

Can a AMD Radeon AI PRO R9700, 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$1,299
Memory32 GB, about 31 GB of it addressable by the GPU at 640 GB/s
Best model it runsQwen3.8 27B — Sonnet-class, 30 tok/s
Pay-back against the APIPays back in 17 years at 500k tokens a day

Run the numbers on this machine

What it runs

ModelSpeedClassGood atMemory
Qwen3.8 27BQ4_K_M 30 tok/s Sonnet-class 19 GB
Qwen3.6 27BQ4_K_M 29 tok/s Haiku-class 19 GB
Qwen3.6 35B-A3BQ4_K_M 132 tok/s Haiku-class 23 GB
Muse Glimmer 30BQ4_K_M 31 tok/s Haiku-class 18 GB
Gemma 4 26B-A4BQ4_K_M 97 tok/s Haiku-class 18 GB
Gemma 4 31B itQ4_K_M 22 tok/s Haiku-class 26 GB
Granite 4.2 30BQ4_K_M 21 tok/s Haiku-class 26 GB
GLM-4.7-FlashQ4_K_M 94 tok/s Haiku-class 20 GB
Gemma 4 12BQ4_K_M 70 tok/s Below every hosted tier 8.0 GB
Qwen3.5 9BQ4_K_M 82 tok/s Below every hosted tier 6.8 GB
Nemotron 3.5 Lightning 30B-A3BQ4_K_M 127 tok/s Below every hosted tier 26 GB
MiniCPM5 2BQ4_K_M 188 tok/s Below every hosted tier 3.0 GB

15 more fit; the calculator lists them all.

The specifics

Chip
Radeon AI PRO R9700 — RDNA 4 · 32 GB GDDR6
Memory bandwidth
640 GB/s
Usable by the GPU
31 GB — Bandwidth: 20 Gbps × 256-bit bus ÷ 8 = 640 GB/s (AMD lists 256-bit and 640 GB/s; XFX and ASUS give 20 Gbps). In llama.cpp, Vulkan (RADV) beat ROCm on decode on this card. 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
300 W (published) — 300 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.
Availability
AMD's launch price, October 2025. Sold through partners, whose prices vary.
Sources
source 1, source 2, source 3, source 4, source 5