AMD Radeon for local LLMs: measured speeds

Every number on this page is a measured generation speed from 13 sourced AMD builds: Radeon AI Pro R9700, RX 7900 XTX, RX 7900 XT and Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S). Each row links the build page and the original benchmark.

Qwen3.6 35B-A3B on each AMD option

Hardware	Best plain decode tok/s	Best with speculative decoding (MTP)	GPU price DE
Radeon AI Pro R9700	117.69 (Q4_K_M, llama.cpp Vulkan)	not measured yet	≈€1,718
Radeon RX 7900 XTX	87.2 (UD-Q4_K_XL, llama-bench 2K, f16 KV, llama.cpp ROCm)	135.2 (UD-Q4_K_XL + MTP (draft-n-max 4), ~16K context, llama.cpp ROCm (llama-server))	–
Radeon RX 7900 XT	not measured yet	not measured yet	≈€869
Strix Halo: Ryzen AI Max+ 395 / Radeon 8060S	62 (Q4_K_M (Unsloth), llama.cpp Vulkan)	93 (Q4_K_M (Unsloth), llama.cpp Vulkan + MTP)	–

Different sources use different quants, context lengths and runs, so compare orders of magnitude, not single digits. For reference, the fastest plain Qwen3.6 35B-A3B results on NVIDIA cards in RigBase: RTX 3090 146-153 tok/s (source build) · RTX 4090 160.6 tok/s (source build) · RTX 5090 258 tok/s (source build).

What the AMD benchmarks agree on

Try llama.cpp's Vulkan backend first. On an R9700, Qwen3.6 27B Q4_K_M ran at 31.9 tok/s on Vulkan vs 27.1 on HIP/ROCm and 26.7 in Ollama (Local AI Frontier). On Strix Halo, the MoE Qwen3.6 35B-A3B Q8_0 ran at 54.22 on Vulkan vs 47.47 on ROCm, while the dense Qwen3.8 27B was a tie (7.77 vs 7.76) (How Quant). The RX 7900 XTX results here are ROCm, so benchmark both backends on your own card.
MTP speculative decoding roughly doubles dense 27B models. R9700: 31.9 → 66.0 tok/s on Qwen3.6 27B. Strix Halo: 9.59 → 18.01 tok/s on Qwen3.8 27B Q6_K (Sleeping Robots). On the RX 7900 XTX, MTP took Qwen3.6 35B-A3B from 79.6 to 135.2 tok/s at ~16K context (gist). Speed-ups depend on how predictable the text is.
Long context costs a lot of speed. One R9700 went from 102.39 tok/s (short) to 44.4 tok/s at 256K on Qwen3.6 35B-A3B (AMDai); the 7900 XTX went from 79.6 at ~16K to 38.3 at ~245K.
Ollama trails llama.cpp on AMD. On Qwen3.8 27B the same R9700 did 30 tok/s in Ollama vs 57 in llama.cpp Vulkan with MTP (Vadi Taslim).
Two R9700s ≠ one 64 GB card. All three dual-R9700 rigs above benchmarked one model per card: one owner reports GPU P2P broken, so every run used one GPU; another measured two single-card copies of Qwen3.6 35B-A3B beating a split server.
Strix Halo is about capacity, not speed. 128 GB unified memory fits models no 24–32 GB card can hold, but dense 27B decode is single-digit without MTP (7.76–9.59 tok/s). On Qwen3.6 35B-A3B one R9700 measured 117.69 tok/s (Q4_K_M) vs 54.22 on Strix Halo (Q8_0, a bigger quant); see Mac vs PC for the other big-memory options.

Radeon AI Pro R9700 (32 GB, RDNA 4)

Price: ≈€1,718 new in Germany (2026-10-16, billiger.de) · 5 sourced build(s)

Model	Quant / setup	Engine	tok/s
Qwen3.6 27B	Q4_K_M + MTP draft head (draft-n-max 3), 64K context	llama.cpp Vulkan	66.0 spec.	build · source
Qwen3.6 27B	Q4_K_M, no MTP	llama.cpp Vulkan	31.9	build · source
Qwen3.6 27B	Q4_K_M	llama.cpp HIP (ROCm)	27.1	build · source
Qwen3.6 27B	Q4_K_M	Ollama (HIP)	26.7	build · source
Qwen3.6 35B-A3B	Q4_K_M	llama.cpp Vulkan	117.69	build · source
Qwen3.6 35B-A3B	Q5_K_M (24.68 GB VRAM)	llama.cpp Vulkan	109.73	build · source
Qwen3.6 35B-A3B	Unsloth UD-Q5_K_XL, short context, f16 KV	llama.cpp Vulkan	102.39	build · source
Qwen3.6 35B-A3B	UD-Q5_K_XL, 32K context	llama.cpp Vulkan	90.52	build · source
Qwen3.6 35B-A3B	UD-Q5_K_XL, 128K context	llama.cpp Vulkan	66.43	build · source
Qwen3.6 35B-A3B	UD-Q5_K_XL, 256K context	llama.cpp Vulkan	44.4	build · source
Qwen3.8 27B	Q4_K_M + MTP speculative decoding, reasoning on	llama.cpp Vulkan	57 spec.	build · source
Qwen3.8 27B	Q4_K_M	Ollama	30	build · source
Qwen3.8 27B	AWQ INT4, single request	vLLM (radiance)	22	build · source
Qwen3.8 Flash-Next	custom int4 experts / 6-bit trunk	strixpro (MTP depth 3 + prompt lookup)	102 spec.	build · source
Qwen3.8 Flash-Next	custom int4 experts / 6-bit trunk	strixpro, ~100K context	82	build · source
Qwen3.8 Flash-Next	custom int4 experts / 6-bit trunk	strixpro, ~980K context	71	build · source
Qwen3.8 Flash-Next	GGUF	llama.cpp (Strix Halo iGPU only)	29-34	build · source
gpt-oss 20B		llama.cpp	163.26	build · source
DeepSeek R1 Distill 14B (older)		llama.cpp	53.72	build · source
Llama 3.2 3B (older)		llama.cpp	195.91	build · source

Radeon RX 7900 XTX (24 GB, RDNA 3)

Price: no dated new-price sample yet · 1 sourced build(s)

Model	Quant / setup	Engine	tok/s
Qwen3.6 35B-A3B	UD-Q4_K_XL + MTP (draft-n-max 4), ~16K context	llama.cpp ROCm (llama-server)	135.2 spec.	build · source
Qwen3.6 35B-A3B	UD-Q4_K_XL + MTP, ~64K context	llama.cpp ROCm (llama-server)	108.2 spec.	build · source
Qwen3.6 35B-A3B	UD-Q4_K_XL, llama-bench 2K, f16 KV	llama.cpp ROCm	87.2	build · source
Qwen3.6 35B-A3B	UD-Q4_K_XL, ~16K real-text context, q4_0 KV	llama.cpp ROCm (llama-server)	79.6	build · source
Qwen3.6 35B-A3B	UD-Q4_K_XL, ~64K context	llama.cpp ROCm (llama-server)	66.0	build · source
Qwen3.6 35B-A3B	UD-Q4_K_XL, ~245K context, q4_0 KV	llama.cpp ROCm (llama-server)	38.3	build · source

Radeon RX 7900 XT (20 GB, RDNA 3)

Price: ≈€869 new in Germany (2026-10-16, billiger.de) · 1 sourced build(s)

Model	Quant / setup	Engine	tok/s
Llama 3.2 3B (older)		llama.cpp	131.2	build · source

Strix Halo: Ryzen AI Max+ 395 / Radeon 8060S (up to 128 GB unified, RDNA 3.5)

Price: sold as a complete mini PC / laptop, no separate GPU price · 6 sourced build(s)

Model	Quant / setup	Engine	tok/s
Gemma 4 26B-A4B	4-bit	Ollama 0.34.0	51.8	build · source
Gemma 4 31B	4-bit	Ollama 0.34.0	8.8	build · source
Qwen3.6 27B	4-bit	Ollama 0.34.0	11.4	build · source
Qwen3.6 35B-A3B	Q4_K_M (Unsloth)	llama.cpp Vulkan + MTP	93 spec.	build · source
Qwen3.6 35B-A3B	Q4_K_M (Unsloth)	llama.cpp Vulkan	62	build · source
Qwen3.6 35B-A3B	Q8_0 + MTP (draft-n-max 3), coding workload	llama.cpp ROCm	60.92 spec.	build · source
Qwen3.6 35B-A3B	Q4_K_M	llama.cpp b11213 Vulkan	59.4	build · source
Qwen3.6 35B-A3B	4-bit	Ollama 0.34.0	54.4	build · source
Qwen3.6 35B-A3B	Q8_0, llama-bench tg128	llama.cpp Vulkan	54.22	build · source
Qwen3.6 35B-A3B	Q8_0	llama.cpp Vulkan	54	build · source
Qwen3.6 35B-A3B	Q4_K_M (Unsloth)	llama.cpp ROCm	51	build · source
Qwen3.6 35B-A3B	llama-bench tg128 (quant per result)	llama.cpp	50.2-62.4	build · source
Qwen3.6 35B-A3B	Q8_0, llama-bench tg128	llama.cpp ROCm	47.47	build · source
Qwen3.8 27B	Q6_K + MTP (draft-n-max 2)	llama.cpp Vulkan (RADV)	18.01 spec.	build · source
Qwen3.8 27B	Q8_0 + MTP (draft-n-max 3), coding workload, reasoning on	llama.cpp ROCm	15.67 spec.	build · source
Qwen3.8 27B	Q6_K + MTP (draft-n-max 5)	llama.cpp Vulkan (RADV)	15.16 spec.	build · source
Qwen3.8 27B	Q8_0 + MTP (draft-n-max 3), coding workload, reasoning on	llama.cpp Vulkan	13.19 spec.	build · source
Qwen3.8 27B	Q4_K_M	llama.cpp/Ollama, ROCm	12.6	build · source
Qwen3.8 27B	Q6_K, no MTP	llama.cpp Vulkan (RADV)	9.59	build · source
Qwen3.8 27B	Q8_0, llama-bench tg128	llama.cpp Vulkan	7.77	build · source
Qwen3.8 27B	Q8_0, llama-bench tg128	llama.cpp ROCm	7.76	build · source
Qwen3.8 27B	Q8_0, --reasoning off	llama.cpp/Ollama, ROCm	7.7	build · source
gpt-oss 120B	MXFP4	llama.cpp b11213 Vulkan	44.7	build · source

Compared with NVIDIA, same author, same prompts

One R9700: 57 tok/s on Qwen3.8 27B (llama.cpp Vulkan + MTP). RTX Pro 6000 Blackwell Max-Q: 150 tok/s (vLLM, NVFP4 + MTP). Both are raw speeds including reasoning tokens, from the same blog post.

Next: check fit with the VRAM calculator, see tok/s per euro (the R9700 and 7900 XT have EU prices), or submit your AMD build; we especially want RX 9070 XT and RX 7900 XTX results on 2026 models.

Not enough VRAM? Rent before you buy. Run the model on a rented 24–96 GB cloud GPU by the hour and see the real speed before spending €1,000+. Try it on Runpod
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.