Which quant of Qwen3.8-27B is fastest on Strix Halo?

IQ4_XS at 14.1 tok/s, ahead of FP4 at 14, Q4_K_M at 12.4, FP8 at 7.7, FP16 at 5. Plain decoding only.

Which backend is fastest for Qwen3.8-27B on Strix Halo?

llama.cpp at 14.1 tok/s, then llama.cpp/rocm at 14, vulkan at 13.9. Each backend figure is its best plain-decoding row and may use a different quant.

How does Strix Halo compare with other boxes for Qwen3.8-27B?

Strix Halo best 36.8 tok/s. Other boxes, best measured: Mac Ultra 42.6, Mac Max 41.5, DGX Spark 23.6. Backends and quants differ, so this is the best known number per box, not a controlled test.

Configs › Qwen3.8-27B › Qwen3.8-27B on Strix Halo: 36.8 tok/s, 19 measured configs

Qwen3.8-27B on Strix Halo: 36.8 tok/s, 19 measured configs

Measured Qwen3.8-27B speeds on Strix Halo: fastest 36.8 tok/s single-stream (FP4, llama.cpp/rocm (speculative decode, mtp-4)), 19 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 Qwen3.8-27B on Strix Halo: 36.8 tok/s (FP4, llama.cpp/rocm (speculative decode, mtp-4)), snapshot 2026-10-29.

Every measured single-stream run of Qwen3.8-27B on Strix Halo

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
36.8	FP4	llama.cpp/rocm (speculative decode, mtp-4)	131072	Verified (ran it)	sypherin	2026-10-29	source
33.8	FP4	speculative decode, mtp-4		Owner-submitted	julianmb	2026-09-16	source
26.9	FP4	speculative decode, mtp	32768	Owner-submitted	julianmb	2026-09-16	source
26.1	Q4_K_XL	gufo (speculative decode, dflash2)		Extracted, source-linked	gufo-org	2026-10-19	source
24.6	UD-Q4_K_XL	llama.cpp/vulkan (speculative decode, mtp)	131072	Verified (ran it)	sypherin	2026-09-22	source
23	UD-Q5_K_XL	llama.cpp (speculative decode, dflash2)		Extracted, source-linked	PieBru	2026-10-14	source
20.4	Q4_K_M	ollama/vulkan (speculative decode, mtp-4)	4096	Structured table	hogeheer	2026-10-14	source
20.4	Q4_K_M	ollama/vulkan (speculative decode, mtp-4)	4096	Structured table	hogeheer	2026-10-29	source
19	FP8	llama.cpp (speculative decode, mtp)		Owner-submitted	julianmb	2026-09-16	source
17	UD-Q6_K_XL	llama.cpp (speculative decode, dflash2)	131072	Extracted, source-linked	PieBru	2026-08-31	source
17	Q8_0	vulkan (speculative decode, mtp)	131072	Owner-submitted	sypherin	2026-09-16	source
14.1	IQ4_XS	llama.cpp	65536	Extracted, source-linked	ilintar	2026-09-06	source
14	FP4	llama.cpp/rocm		Extracted, source-linked	julianmb	2026-09-06	source
13.9	FP4	vulkan	16384	Owner-submitted	julianmb	2026-09-10	source
12.4	Q4_K_M			Owner-submitted	julianmb	2026-09-16	source
12.3	Q4_K_M	llama.cpp	32768	Owner-submitted	julianmb	2026-09-16	source
12.3	Q4_K_M	llama.cpp/rocm		Extracted, source-linked	julianmb	2026-09-06	source
7.7	FP8	llama.cpp/rocm		Extracted, source-linked	julianmb	2026-09-06	source
5	FP16	llama.cpp		Owner-submitted	julianmb	2026-09-16	source

Best quant on Strix Halo

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

IQ4_XS: 14.1 tok/s, fastest (llama.cpp)
FP4: 14 tok/s, 1% lower (llama.cpp/rocm)
Q4_K_M: 12.4 tok/s, 12% lower ()
FP8: 7.7 tok/s, 45% lower (llama.cpp/rocm)
FP16: 5 tok/s, 65% lower (llama.cpp)

Backends on Strix Halo

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

llama.cpp: 14.1 tok/s (IQ4_XS)
llama.cpp/rocm: 14 tok/s (FP4)
vulkan: 13.9 tok/s (FP4)

Qwen3.8-27B on other boxes

Mac Ultra: 42.6 tok/s best of 1 config
Mac Max: 41.5 tok/s best of 4 configs
DGX Spark: 23.6 tok/s best of 2 configs

Questions and answers

How fast is Qwen3.8-27B on Strix Halo?: The fastest of 19 measured configs is 36.8 tok/s single-stream (FP4, llama.cpp/rocm (speculative decode, mtp-4)). The slowest is 5 tok/s.
Which quant of Qwen3.8-27B is fastest on Strix Halo?: IQ4_XS at 14.1 tok/s, ahead of FP4 at 14, Q4_K_M at 12.4, FP8 at 7.7, FP16 at 5. Plain decoding only.
Which backend is fastest for Qwen3.8-27B on Strix Halo?: llama.cpp at 14.1 tok/s, then llama.cpp/rocm at 14, vulkan at 13.9. Each backend figure is its best plain-decoding row and may use a different quant.
How does Strix Halo compare with other boxes for Qwen3.8-27B?: Strix Halo best 36.8 tok/s. Other boxes, best measured: Mac Ultra 42.6, Mac Max 41.5, DGX Spark 23.6. Backends and quants differ, so this is the best known number per box, not a controlled test.

See also: Qwen3.8-27B 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.