AMD · Workstation · RDNA 4 (Navi 48, gfx1201)

AMD Radeon AI PRO R9700

AMD lists cooling as 'Active'; partner boards are rear-exhaust dual-slot designs. 12V-2x6 power connector.

VRAM

32 GB

GDDR6 ECC (Linux)

Memory bandwidth

640 GB/s

FP16 compute

191 TFLOPS

dense tensor

Price

~$1,299

new, maker or suggested price

What fits at 4-bit

The 4-bit file plus 8,192 tokens of context and 1.5 GB of overhead, against the VRAM of 1, 2, 4 and 6 cards. Speeds are measured medians where someone has published a run at that card count; a blank speed means no run, not a slow one. Machine pages carry estimates for the gaps.

Model	Needs	1 card (32 GB)	2 cards (64 GB)	4 cards (128 GB)	6 cards (192 GB)
Qwen3.5 4B4B · Q4_K_M	4.3 GB	fits	fits	fits	fits
Llama 3.1 8B Instruct8B · Q4_K_M	7.1 GB	fits	fits	fits	fits
Qwen3 8B8.2B · Q4_K_M	7.3 GB	fits	fits	fits	fits
Gemma 4 E4B IT8B · Q4_K_M	6.5 GB	fits	fits	fits	fits
Granite 4.2 8B8.8B · Q4_K_M	7.9 GB	fits	fits	fits	fits
Qwen3.5 9B9B · Q4_K_M	7.0 GB	fits	fits	fits	fits
Gemma 4 12B IT11.95B · Q4_K_M	8.2 GB	fits	fits	fits	fits
Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M	10.4 GB	fits	fits	fits	fits
Phi-414.7B · Q4_K_M	11.3 GB	fits	fits	fits	fits
gpt-oss-20b21B, 3.6B active · MXFP4 (native)	12.5 GB	142 tok/smedian of 1 run	fits	fits	fits
Devstral Small 2 24B (2512)24B · Q4_K_M	16.1 GB	fits	fits	fits	fits
Qwen3.6 27B27B · Q4_K_M	17.7 GB	24 tok/smedian of 2 runs	fits	fits	fits
Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M	17.4 GB	11 tok/smedian of 1 run	fits	fits	fits
Qwen3.8 27B27B · Q4_K_M	18.2 GB	30 tok/smedian of 2 runs	fits	fits	fits
Granite 4.2 30B29.3B · Q4_K_M	20.3 GB	fits	fits	fits	fits
Gemma 4 31B IT30.7B · Q4_K_M	18.9 GB	fits	fits	fits	fits
Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M	19.5 GB	fits	fits	fits	fits
Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M	21.9 GB	fits	fits	fits	fits
Qwen3.6 35B-A3B35B, 3B active · Q4_K_M	22.4 GB	fits	fits	fits	fits
Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M	25.3 GB	fits	fits	fits	fits
Llama 3.3 70B Instruct70B · Q4_K_M	43.6 GB	too large	fits	fits	fits
Qwen3 Coder Next80B, 3B active · Q4_K_M	46.9 GB	too large	fits	fits	fits
gpt-oss-120b117B, 5.1B active · MXFP4 (native)	60.2 GB	too large	fits	fits	fits
Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M	62.4 GB	too large	fits	fits	fits
Mistral Small 4 119B (2603)119B, 6.5B active · Q4_K_M	69.3 GB	too large	too large	fits	fits
GLM-4.5-Air106B, 12B active · Q4_K_M	70.9 GB	too large	too large	fits	fits
Mistral Medium 3.5 128B128B · Q4_K_M	74.0 GB	too large	too large	fits	fits
Qwen3.5 122B-A10B122B, 10B active · Q4_K_M	73.0 GB	too large	too large	fits	fits
Nemotron 3 Super 120B-A12B120B, 12B active · Q4_K_M	82.6 GB	too large	too large	fits	fits
Qwen3.8 Flash-Next125B, 6B active · Q4_K_M	113.1 GB	too large	too large	fits	fits
MiniMax M2.7228.7B · Q4_K_M	132.7 GB	too large	too large	too large	fits
Hy3295B, 21B active · Q4_K_M	173.6 GB	too large	too large	too large	fits
GLM-5.3 Flash320B, 18B active · UD-Q4_K_XL	187.6 GB	too large	too large	too large	fits

Machines built on it

tiny corp tinybox red flash~$22,000

Measured speeds

One card

Single-user runs on 1 card
Model	Quant	Backend	Generation	Prompt	Source
Gemma 4 26B-A4B IT	UD-Q4_K_M	llama.cpp b8757 (Vulkan, RADV)	11 tok/s	209 tok/s512-token prompt	owner report2026-04
gpt-oss-20b	MXFP4 MoE	llama.cpp b8348 ROCm	142 tok/s	2,178 tok/s2,048-token prompt	llama.cpp benchmark discussion2026-03
Llama 3.1 8B Instruct	BF16	vLLM v0.20.2 (NCCL_PROTO=Simple)	39 tok/s	–	owner report2026-05
Qwen3.6 27B	UD-Q4_K_XL	llama.cpp b9203 (Vulkan RADV Mesa 26.1.0)	28 tok/s	924 tok/s2,048-token prompt	llama.cpp benchmark discussion2026-05
Qwen3.6 27B	Q4_K_M	llama.cpp (Vulkan RADV Mesa 26.0.3, build not stated)	20 tok/s	470 tok/s945-token prompt	llama.cpp benchmark discussion2026-05
Qwen3.6 27B	UD-Q4_K_XL	llama.cpp b9203 (Vulkan RADV Mesa 26.1.0)	–	826 tok/s16,384-token prompt	llama.cpp benchmark discussion2026-05
Qwen3.8 27B	UD-Q4_K_XL	llama.cpp (commit 9cc3394, Vulkan, AMD proprietary Windows driver)	31 tok/s	719 tok/s512-token prompt	llama.cpp benchmark discussion2026-09
Qwen3.8 27B	UD-Q4_K_XL	llama.cpp (Lemonade llamacpp-rocm b1322, HIP gfx1201, Windows)	29 tok/s	1,106 tok/s512-token prompt	llama.cpp benchmark discussion2026-09

2 cards

Single-user runs on 2 cards
Model	Quant	Backend	Generation	Prompt	Context	Source
Qwen3.6 27B	UD-Q8_K_XL	llama.cpp (ROCm 7.2.0, build not stated)	16 tok/s	355 tok/s3,243-token prompt	3,243	llama.cpp benchmark discussion2026-05
Qwen3.6 35B-A3B	UD-Q8_K_XL	llama.cpp b9203 (Vulkan RADV Mesa 26.1.0)	73 tok/s	3,196 tok/s2,048-token prompt	0	llama.cpp benchmark discussion2026-05
Qwen3.6 35B-A3B	UD-Q8_K_XL	llama.cpp b9203 (Vulkan RADV Mesa 26.1.0)	–	3,249 tok/s65,536-token prompt	0	llama.cpp benchmark discussion2026-05
Qwen3.8 Flash-Next	UD-IQ3_XXS	llama.cpp b10715 (unsloth build, Vulkan)	46 tok/s	894 tok/s2,048-token prompt	0	llama.cpp benchmark discussion2026-09
Qwen3.8 Flash-Next	UD-IQ3_XXS	llama.cpp fork (JohnTDI-cpu/llama.cpp-flash-next-rdna4, ROCm 7.2.4 HIP)	40 tok/s	942 tok/s2,048-token prompt	0	llama.cpp benchmark discussion2026-09

4 cards

Single-user runs on 4 cards
Model	Quant	Backend	Generation	Prompt	Context	Source
Qwen3.6 27B	UD-Q8_K_XL	llama.cpp (ROCm 7.2.0, build not stated)	15 tok/s	329 tok/s3,243-token prompt	3,243	llama.cpp benchmark discussion2026-05

Speculative decoding, several users, or long context

Kept apart because each measures a different workload from one person chatting near the start of a conversation.

Other runs on this card
Model	Quant	Backend	Generation	Prompt	Context	Measured on	Source
Gemma 4 26B-A4B IT	BF16	vLLM v0.20.2 (NCCL_PROTO=Simple)	26 tok/s	–	0	4× Radeon AI PRO R9700	owner report2026-05
gpt-oss-120b	BF16 (no quantization)	vLLM v0.20.2 (NCCL_PROTO=Simple)	5.9 tok/s	–	0	4× Radeon AI PRO R9700	owner report2026-05
Llama 3.1 8B Instruct	BF16	vLLM v0.20.2 (NCCL_PROTO=Simple)	68 tok/s	–	0	2× Radeon AI PRO R9700	owner report2026-05
Llama 3.1 8B Instruct	BF16	vLLM v0.20.2 (NCCL_PROTO=Simple)	64 tok/s	–	0	4× Radeon AI PRO R9700	owner report2026-05
Qwen3.5 122B-A10B	UD-Q6_K_XL	llama.cpp b89xx (stated 'b89+', ROCm 7.2.0)256 requests at once	32 tok/s	230 tok/s	0	4× Radeon AI PRO R9700	llama.cpp benchmark discussion2026-04
Qwen3.5 122B-A10B	GPTQ Int4	vLLM v0.17.0-dev + AITER	54 tok/s	–	0	4× Radeon AI PRO R9700	owner report2026-05
Qwen3.6 27B	BF16	vLLM v0.20.2 (NCCL_PROTO=Simple)	17 tok/s	–	0	2× Radeon AI PRO R9700	owner report2026-05
Qwen3.6 27B	UD-Q4_K_XL	llama.cpp b9203 (Vulkan RADV Mesa 26.1.0)	52 tok/s	–	0	Radeon AI PRO R9700	llama.cpp benchmark discussion2026-05
Qwen3.6 27B	Q4_K_M	llama.cpp (Vulkan RADV Mesa 26.0.3, build not stated)	44 tok/s	–	0	Radeon AI PRO R9700	llama.cpp benchmark discussion2026-05
Qwen3.6 27B	Q6_K	llama.cpp fork (b9940 + JohnTDI-cpu HIP Direct-P2P AllReduce)	37 tok/s	820 tok/s512-token prompt	0	2× Radeon AI PRO R9700	llama.cpp benchmark discussion2026-08
Qwen3.6 27B	Q6_K	llama.cpp b9940 (ROCm 7.2.0 HIP)	34 tok/s	812 tok/s512-token prompt	0	2× Radeon AI PRO R9700	llama.cpp benchmark discussion2026-08
Qwen3.6 27B	Q8_0	llama.cpp b9209 (ghcr full-rocm image, ROCm 7.2.1)	43 tok/s	960 tok/s3,666-token prompt	3,666	Radeon AI PRO R9700	llama.cpp benchmark discussion2026-05
Qwen3.6 27B	Q8_0	llama.cpp b9209 (ghcr full-rocm image, ROCm 7.2.1)	36 tok/s	898 tok/s22,108-token prompt	22,108	Radeon AI PRO R9700	llama.cpp benchmark discussion2026-05
Qwen3.6 27B	Q8_0	llama.cpp b9209 (ghcr full-rocm image, ROCm 7.2.1)	30 tok/s	756 tok/s47,679-token prompt	47,679	Radeon AI PRO R9700	llama.cpp benchmark discussion2026-05
Qwen3.6 27B	Q8_0	llama.cpp b9209 (ghcr full-rocm image, ROCm 7.2.1)	28 tok/s	708 tok/s58,723-token prompt	58,723	Radeon AI PRO R9700	llama.cpp benchmark discussion2026-05
Qwen3.6 35B-A3B	FP8	vLLM v0.20.2 (NCCL_PROTO=Simple)	17 tok/s	–	0	2× Radeon AI PRO R9700	owner report2026-05
Qwen3.6 35B-A3B	FP8	vLLM v0.20.2 (NCCL_PROTO=Simple)	16 tok/s	–	0	4× Radeon AI PRO R9700	owner report2026-05
Qwen3.8 27B	MXFP4	Radiance v1.2.3 (C++/HIP, ROCm in container)	133 tok/s	–	0	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	MXFP4	Radiance (version not stated; llama-benchy client)	131 tok/s	4,219 tok/s4,096-token prompt	65,536	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	MXFP4	Radiance (version not stated; llama-benchy client)	128 tok/s	4,863 tok/s4,096-token prompt	0	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	MXFP4	Radiance (version not stated; llama-benchy client)	125 tok/s	4,990 tok/s4,096-token prompt	16,384	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	MXFP4	Radiance (version not stated; llama-benchy client)	88 tok/s	3,379 tok/s4,096-token prompt	16,384	Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	MXFP4	Radiance (version not stated; llama-benchy client)	78 tok/s	2,699 tok/s4,096-token prompt	65,536	Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	MXFP4	Radiance (version not stated; llama-benchy client)	76 tok/s	3,451 tok/s4,096-token prompt	0	Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	UD-Q4_K_XL	llama.cpp (commit 9cc3394, Vulkan, AMD proprietary Windows driver)	55 tok/s	–	0	Radeon AI PRO R9700	llama.cpp benchmark discussion2026-09
Qwen3.8 27B	Q6_K	llama.cpp fork (b10434 + JohnTDI-cpu HIP Direct-P2P + Phase 13 Q6_K F32 cast)	37 tok/s	1,019 tok/s512-token prompt	0	2× Radeon AI PRO R9700	llama.cpp benchmark discussion2026-08
Qwen3.8 27B	Q6_K	llama.cpp b10434 (ROCm/HIP 7.2)	34 tok/s	878 tok/s512-token prompt	0	2× Radeon AI PRO R9700	llama.cpp benchmark discussion2026-08
Qwen3.8 27B	FP8	Radiance v1.2.3 (C++/HIP, ROCm in container)	169 tok/s	–	0	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	FP8	Radiance v1.2.3 (C++/HIP, ROCm in container)	130 tok/s	–	0	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	FP8	Radiance v1.2.3 (C++/HIP, ROCm in container)	111 tok/s	–	0	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	Q8_0	llama.cpp (commit 0d9ceae1e, Vulkan RADV 25.2.8)	65 tok/s	–	0	Radeon AI PRO R9700	llama.cpp benchmark discussion2026-08
Qwen3.8 27B	FP8	Radiance v1.2.3 (C++/HIP, ROCm in container)	23 tok/s	–	2,048	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 27B	FP8	Radiance v1.2.3 (C++/HIP, ROCm in container)	1.7 tok/s	–	200,000	2× Radeon AI PRO R9700	owner report2026-10
Qwen3.8 Flash-Next	UD-IQ3_XXS	llama.cpp fork (JohnTDI-cpu/llama.cpp-flash-next-rdna4, ROCm 7.2.4 HIP)	62 tok/s	820 tok/s2,048-token prompt	0	2× Radeon AI PRO R9700	llama.cpp benchmark discussion2026-09
Qwen3.8 Flash-Next	UD-IQ3_XXS	llama.cpp b10715 (unsloth build, Vulkan)	50 tok/s	690 tok/s2,048-token prompt	0	2× Radeon AI PRO R9700	llama.cpp benchmark discussion2026-09

Specifications

Memory: 32 GB GDDR6 ECC (Linux)
FP16 basis: AMD 'Peak Half Precision (FP16 Matrix)' dense; 383 with structured sparsity.
Board power: 300 W
Slots: 2
PCIe: Gen 5 x16
NVLink: No
Cooling: Blower
Software: ROCm, Vulkan
Released: 2025
Availability: sold new
Price basis: AMD suggested price at the Oct 2025 retail launch; Oct 2026 street listings were reported well above it.
Maker page: amd.com

Sources

amd.com: 32 GB GDDR6 256-bit, 640 GB/s, ECC Linux only, 300 W TBP, PCIe 5.0 x16, double slot, 191/383 TFLOPS FP16 matrix
guru3d.com: $1,299 suggested price
rocm.docs.amd.com: ROCm 7.14 supported-GPU list: 7900 XTX/W7900 (gfx1100), 9070 XT/R9700 (gfx1201), MI100 (gfx908) supported; MI50 (gfx906) not supported; Radeon RX cards Ubuntu-only