Which quant of Qwen3.8-Flash-Next is fastest on Strix Halo?

NVFP4 at 41.3 tok/s, ahead of W4B at 37.6, IQ4_XS at 27.2, Q4_K_XL at 24, IQ3_XXS at 23.5. Plain decoding only.

Which backend is fastest for Qwen3.8-Flash-Next on Strix Halo?

halogen/rocm at 37.6 tok/s, then llama.cpp/vulkan at 27.2, llama.cpp at 23.5. 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-Flash-Next?

Strix Halo best 63.7 tok/s. Other boxes, best measured: Mac Ultra 76.4, Mac Max 57.5, DGX Spark 44.2. Backends and quants differ, so this is the best known number per box, not a controlled test.

Configs › Qwen3.8-Flash-Next › Qwen3.8-Flash-Next on Strix Halo: 63.7 tok/s, 15 measured configs

Qwen3.8-Flash-Next on Strix Halo: 63.7 tok/s, 15 measured configs

Measured Qwen3.8-Flash-Next speeds on Strix Halo: fastest 63.7 tok/s single-stream (UD-IQ4_XS, halogen/rocm (speculative decode, mtp)), 15 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-Flash-Next on Strix Halo: 63.7 tok/s (UD-IQ4_XS, halogen/rocm (speculative decode, mtp)), snapshot 2026-10-29.

Every measured single-stream run of Qwen3.8-Flash-Next on Strix Halo

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
63.7	UD-IQ4_XS	halogen/rocm (speculative decode, mtp)	131072	Verified (ran it)	sypherin	2026-10-29	source
57.5		crucible (speculative decode, mtp)	32768	Extracted, source-linked	jaedync	2026-10-25	source
55.3		llama.cpp/hip (speculative decode, mtp)	262144	Extracted, source-linked	abliter8-ai	2026-09-18	source
48.5		strixite/hip (speculative decode, mtp)	512000	Extracted, source-linked	shawnshekari	2026-10-04	source
48.1		exllamav3 (speculative decode, spec)	500000	Extracted, source-linked	Yamz-Labs	2026-10-17	source
46	W4B	halogen/rocm (speculative decode, mtp)	32768	Extracted, source-linked	peonist-ai	2026-09-30	source
43.9	UD-Q4_K_XL	llama.cpp/rocm (speculative decode, mtp)		Extracted, source-linked	kyuz0	2026-10-14	source
41.3	NVFP4			Extracted, source-linked	Soulmate-Halo	2026-10-14	source
40.8	UD-Q4_K_XL	gufo (speculative decode, mtp)	65536	Extracted, source-linked	huppiflupp	2026-10-24	source
37.6	W4B	halogen/rocm (baseline)	1500	Extracted, source-linked	peonist-ai	2026-09-30	source
35.8	UD-Q4_K_XL	gufo (speculative decode, mtp)		Extracted, source-linked	gufo-org	2026-10-19	source
35.7	IQ3_XXS	llama.cpp (speculative decode, mtp)	156000	Extracted, source-linked	EasiiX	2026-08-29	source
27.2	IQ4_XS	llama.cpp/vulkan		Structured table	hogeheer	2026-10-29	source
24	Q4_K_XL	llama.cpp/vulkan		Extracted, source-linked	routhjim	2026-08-29	source
23.5	IQ3_XXS	llama.cpp (baseline)	156000	Extracted, source-linked	EasiiX	2026-09-01	source

Best quant on Strix Halo

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

NVFP4: 41.3 tok/s, fastest ()
W4B: 37.6 tok/s, 9% lower (halogen/rocm (baseline))
IQ4_XS: 27.2 tok/s, 34% lower (llama.cpp/vulkan)
Q4_K_XL: 24 tok/s, 42% lower (llama.cpp/vulkan)
IQ3_XXS: 23.5 tok/s, 43% lower (llama.cpp (baseline))

Backends on Strix Halo

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

halogen/rocm: 37.6 tok/s (W4B)
llama.cpp/vulkan: 27.2 tok/s (IQ4_XS)
llama.cpp: 23.5 tok/s (IQ3_XXS)

Qwen3.8-Flash-Next on other boxes

Mac Ultra: 76.4 tok/s best of 1 config
Mac Max: 57.5 tok/s best of 2 configs
DGX Spark: 44.2 tok/s best of 8 configs

Questions and answers

How fast is Qwen3.8-Flash-Next on Strix Halo?: The fastest of 15 measured configs is 63.7 tok/s single-stream (UD-IQ4_XS, halogen/rocm (speculative decode, mtp)). The slowest is 23.5 tok/s.
Which quant of Qwen3.8-Flash-Next is fastest on Strix Halo?: NVFP4 at 41.3 tok/s, ahead of W4B at 37.6, IQ4_XS at 27.2, Q4_K_XL at 24, IQ3_XXS at 23.5. Plain decoding only.
Which backend is fastest for Qwen3.8-Flash-Next on Strix Halo?: halogen/rocm at 37.6 tok/s, then llama.cpp/vulkan at 27.2, llama.cpp at 23.5. 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-Flash-Next?: Strix Halo best 63.7 tok/s. Other boxes, best measured: Mac Ultra 76.4, Mac Max 57.5, DGX Spark 44.2. Backends and quants differ, so this is the best known number per box, not a controlled test.

See also: Qwen3.8-Flash-Next 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.