Measured review · llama.cpp · 256K context
Radeon AI PRO R9700 review for local LLMs
If you want the best local model that fits on one AMD card at full context, the R9700 is the card to buy: 32 GB holds Qwen3.6-35B-A3B at Q5 with the whole 262K context, and it decodes at about 100 tok/s on short prompts on the local desktop.
As an Amazon Associate I earn from qualifying purchases. Some Newegg links use Rakuten Advertising and some B&H links use B&H’s affiliate program when those IDs are configured. Micro Center and AMD Store links are plain “also available at” links — no tracking. Buttons say “Check price” and never invent or live-quote a price. Full disclosure.
The short answer
Do not buy it for gaming, do not pay a marketplace markup without checking the seller, and do not buy a second one expecting the 35B to get faster. Two cards are for running two copies at once or for big Flash-class models with DDR5 offload. Every number below is from our own run cards on the local desktop (llama.cpp Vulkan).
What it runs (measured)
One R9700, Qwen3.6-35B-A3B, llama.cpp Vulkan, f16 KV, local desktop. Decode (prefill stays on the run cards). The site default, UD-Q5_K_XL, fits at 262K on one card with about 3 GB to spare (28.7 GB peak). Long context costs speed: decode falls from about 102 tok/s at a short prompt to about 44 tok/s with 262K tokens already in the cache. Prefill is about 2,700 tok/s short and about 520 tok/s at 262K depth — agent turns that resend a big prompt wait on this number.
| Quant | 4K | 32K | 128K | 262K |
|---|---|---|---|---|
| Q4_K_M | 117.7 | 101.9 | 72.5 | 47.8 |
| UD-Q5_K_XL | 102.4 | 90.5 | 66.4 | 44.4 |
| Q6_K | 103.2 | 71.2 | 54.5 | 37.8 |
Does a second R9700 help?
Not for one 35B chat: splitting it across two cards is slower at every depth we measured (Q4_K_M 117.7 vs 93.3 at 4K, 47.8 vs 42.9 at 262K; UD-Q5_K_XL 102.4 vs 82.8, 44.4 vs 42.0). A second card earns its place two ways: two independent 35B copies (172–177 tok/s combined) or a much bigger model (Qwen 3.8 Flash UD-Q6_K_XL, two cards plus DDR5) at 32.6 tok/s on Strata vs 16 on llama.cpp at 262K. “35B on two cards” means one full copy per card.
| Quant × context | One card | Split across two |
|---|---|---|
| Q4_K_M · 4K | 117.7 | 93.3 |
| Q4_K_M · 262K | 47.8 | 42.9 |
| UD-Q5_K_XL · 4K | 102.4 | 82.8 |
| UD-Q5_K_XL · 262K | 44.4 | 42.0 |
The downsides, plainly
Price: AMD’s launch MSRP is $1,299. We omit today’s street ask. If the only offer is a big marketplace markup, wait. Blower noise and heat: all board-partner cards are 2-slot blowers. Software lags NVIDIA in places — llama.cpp’s newest MoE prefill fusion and sparse flash attention for Flash-class models are on for NVIDIA (and AMD Instinct) but not for RDNA4 yet. Windows is second-class; Linux with ROCm or Vulkan is the supported path. It works as a display card, but the desktop takes some VRAM (about 0.5 GB on the local desktop) that the 262K fit needs.
Who should buy one
| You have | Our call |
|---|---|
| A 24–32 GB Radeon or NVIDIA card already | Try the 35B on it first. Buy nothing yet. |
| A 16 GB card and want the 35B at full speed and full context | One R9700. |
| One R9700 and want it faster | A second card will not speed up the 35B. Buy it only for two users or for Flash with DDR5. |
| A gaming-first build | Not this card. |
Board-partner cards in one place
ASRock Creator, ASUS Turbo, Gigabyte AI TOP, Sapphire, and XFX all ship 2-slot blowers. Length, power connector, and outputs live on the R9700 sheet. We do not invent a temperature or noise figure that is not already cited there.
Where to buy
Check current stock. AMD launch MSRP is $1,299. Marketplace listings use that seller’s return policy. We may earn a commission if you buy through a retailer hop; it never changes the verdict.
FAQ
- Is the R9700 good for local LLMs?
- Yes for 32B–35B-class models at full context on one card: measured 102 tok/s (UD-Q5_K_XL, short prompt) and 44 tok/s at 262K on the local desktop.
- This review
- How fast is the R9700 in llama.cpp?
- 4K / 262K decode: Q4_K_M 118 / 48, UD-Q5_K_XL 102 / 44, Q6_K 103 / 38 tok/s (Vulkan, one card, local desktop).
- Benchmarks
- Is two R9700s faster?
- Not for one 35B. Yes for two concurrent copies (172–177 tok/s combined) or for bigger models with DDR5 offload.
- One vs two GPUs
Sources
- S-RESULTS — Official amdai-bench/run@1 cards — our lab · checked 2026-11-02 · Local desktop llama.cpp Vulkan sweep
- S-COMPARE — One GPU vs two — 35B matrix — our lab · checked 2026-11-02 · Rendered from the same run cards
- S-METHOD — How we test — our lab · checked 2026-11-02 · amdai-bench/run@1
How we test — Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.