Which quant of Qwen3.6-27B is fastest on Strix Halo?

Q8_0 at 7.8 tok/s, ahead of UD-Q8_K_XL at 6.6. Plain decoding only.

Which backend is fastest for Qwen3.6-27B on Strix Halo?

llama.cpp/rocm-7.2.4 at 7.8 tok/s, then llama.cpp/rocm-7.14 at 6.6, llama.cpp/vulkan-radv at 6.6, llama.cpp/vulkan-radv-performance at 6.6, llama.cpp/rocm at 6.5, llama.cpp/rocm-7.2.3-mtp at 6.3, llama.cpp/vulkan at 6.3, llama.cpp/vulkan-radv-mtp at 6.1. Each backend figure is its best plain-decoding row and may use a different quant.

How does Strix Halo compare with other boxes for Qwen3.6-27B?

Strix Halo best 21 tok/s. Other boxes, best measured: DGX Spark 37.3. Backends and quants differ, so this is the best known number per box, not a controlled test.

Configs › Qwen3.6-27B › Qwen3.6-27B on Strix Halo: 21 tok/s, 18 measured configs

Qwen3.6-27B on Strix Halo: 21 tok/s, 18 measured configs

Measured Qwen3.6-27B speeds on Strix Halo: fastest 21 tok/s single-stream (UD-Q4_K_XL, llama.cpp (speculative decode, mtp-5)), 18 configs with quant, backend and context, each linked to its source, snapshot 2026-10-29.

Snapshot 2026-10-29 · 413 configs on AI boxes and 288 on GPU cards · 121 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.

Fastest measured single-stream run of Qwen3.6-27B on Strix Halo: 21 tok/s (UD-Q4_K_XL, llama.cpp (speculative decode, mtp-5)), snapshot 2026-10-29.

Every measured single-stream run of Qwen3.6-27B on Strix Halo

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
21	UD-Q4_K_XL	llama.cpp (speculative decode, mtp-5)	262144	Owner-submitted	sypherin	2026-09-16	source
13.5	UD-Q8_K_XL	llama.cpp/rocm (speculative decode, mtp-3)		Structured table	kyuz0	2026-09-06	source
13.3	UD-Q8_K_XL	llama.cpp/vulkan (speculative decode, mtp-3)		Structured table	kyuz0	2026-09-06	source
12.8	UD-Q8_K_XL	llama.cpp/rocm-7.2.3-mtp (speculative decode, mtp-3)		Structured table	kyuz0	2026-10-29	source
12.4	UD-Q8_K_XL	llama.cpp/rocm (speculative decode, mtp-2)		Structured table	kyuz0	2026-09-06	source
12.2	UD-Q8_K_XL	llama.cpp/vulkan-radv-mtp (speculative decode, mtp-3)		Structured table	kyuz0	2026-10-29	source
11.7	UD-Q8_K_XL	llama.cpp/rocm-7.2.3-mtp (speculative decode, mtp-2)		Structured table	kyuz0	2026-10-29	source
11.7	UD-Q8_K_XL	llama.cpp/vulkan (speculative decode, mtp-2)		Structured table	kyuz0	2026-09-06	source
10.8	UD-Q8_K_XL	llama.cpp/vulkan-radv-mtp (speculative decode, mtp-2)		Structured table	kyuz0	2026-10-29	source
7.8	Q8_0	llama.cpp/rocm-7.2.4		Structured table	kyuz0	2026-10-29	source
6.6	UD-Q8_K_XL	llama.cpp/rocm-7.14		Structured table	kyuz0	2026-10-29	source
6.6	UD-Q8_K_XL	llama.cpp/rocm-7.2.4		Structured table	kyuz0	2026-10-29	source
6.6	UD-Q8_K_XL	llama.cpp/vulkan-radv		Structured table	kyuz0	2026-10-29	source
6.6	UD-Q8_K_XL	llama.cpp/vulkan-radv-performance		Structured table	kyuz0	2026-10-29	source
6.5	UD-Q8_K_XL	llama.cpp/rocm (baseline)		Structured table	kyuz0	2026-09-06	source
6.3	UD-Q8_K_XL	llama.cpp/rocm-7.2.3-mtp (baseline)		Structured table	kyuz0	2026-10-29	source
6.3	UD-Q8_K_XL	llama.cpp/vulkan (baseline)		Structured table	kyuz0	2026-09-06	source
6.1	UD-Q8_K_XL	llama.cpp/vulkan-radv-mtp (baseline)		Structured table	kyuz0	2026-10-29	source

Best quant on Strix Halo

Fastest plain-decoding run per quant, versus the fastest quant.

Q8_0: 7.8 tok/s, fastest (llama.cpp/rocm-7.2.4)
UD-Q8_K_XL: 6.6 tok/s, 15% lower (llama.cpp/rocm-7.14)

Backends on Strix Halo

Fastest plain-decoding run per backend. Quants can differ between rows.

llama.cpp/rocm-7.2.4: 7.8 tok/s (Q8_0)
llama.cpp/rocm-7.14: 6.6 tok/s (UD-Q8_K_XL)
llama.cpp/vulkan-radv: 6.6 tok/s (UD-Q8_K_XL)
llama.cpp/vulkan-radv-performance: 6.6 tok/s (UD-Q8_K_XL)
llama.cpp/rocm: 6.5 tok/s (UD-Q8_K_XL)
llama.cpp/rocm-7.2.3-mtp: 6.3 tok/s (UD-Q8_K_XL)
llama.cpp/vulkan: 6.3 tok/s (UD-Q8_K_XL)
llama.cpp/vulkan-radv-mtp: 6.1 tok/s (UD-Q8_K_XL)

Qwen3.6-27B on other boxes

DGX Spark: 37.3 tok/s best of 9 configs

Questions and answers

How fast is Qwen3.6-27B on Strix Halo?: The fastest of 18 measured configs is 21 tok/s single-stream (UD-Q4_K_XL, llama.cpp (speculative decode, mtp-5)). The slowest is 6.1 tok/s.
Which quant of Qwen3.6-27B is fastest on Strix Halo?: Q8_0 at 7.8 tok/s, ahead of UD-Q8_K_XL at 6.6. Plain decoding only.
Which backend is fastest for Qwen3.6-27B on Strix Halo?: llama.cpp/rocm-7.2.4 at 7.8 tok/s, then llama.cpp/rocm-7.14 at 6.6, llama.cpp/vulkan-radv at 6.6, llama.cpp/vulkan-radv-performance at 6.6, llama.cpp/rocm at 6.5, llama.cpp/rocm-7.2.3-mtp at 6.3, llama.cpp/vulkan at 6.3, llama.cpp/vulkan-radv-mtp at 6.1. Each backend figure is its best plain-decoding row and may use a different quant.
How does Strix Halo compare with other boxes for Qwen3.6-27B?: Strix Halo best 21 tok/s. Other boxes, best measured: DGX Spark 37.3. Backends and quants differ, so this is the best known number per box, not a controlled test.

See also: Qwen3.6-27B on every box · Strix Halo hardware page (from S$4,329 in Singapore today).

Last updated: snapshot 2026-10-29; newest source check 2026-10-29.

More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.