Which quant of DeepSeek-V4-Flash-0731 is fastest on Strix Halo?

UD-IQ3_XXS at 19.8 tok/s, ahead of UD-IQ2_XXS at 16.2. Plain decoding only.

Which backend is fastest for DeepSeek-V4-Flash-0731 on Strix Halo?

llama.cpp/vulkan-radv-performance at 19.8 tok/s, then llama.cpp/rocm-7.14 at 16.2, llama.cpp/rocm-7.2.4 at 16.1, llama.cpp/rocm-7.14-pr26592 at 14.8, vulkan at 12, llama.cpp/vulkan-radv at 9.1. Each backend figure is its best plain-decoding row and may use a different quant.

How does Strix Halo compare with other boxes for DeepSeek-V4-Flash-0731?

Strix Halo best 35.7 tok/s. Other boxes, best measured: DGX Spark 32.8. Backends and quants differ, so this is the best known number per box, not a controlled test.

Configs › DeepSeek-V4-Flash-0731 › DeepSeek-V4-Flash-0731 on Strix Halo: 35.7 tok/s, 14 measured configs

DeepSeek-V4-Flash-0731 on Strix Halo: 35.7 tok/s, 14 measured configs

Measured DeepSeek-V4-Flash-0731 speeds on Strix Halo: fastest 35.7 tok/s single-stream (UD-IQ3_XXS, llama.cpp/vulkan (speculative decode, dspark)), 14 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 DeepSeek-V4-Flash-0731 on Strix Halo: 35.7 tok/s (UD-IQ3_XXS, llama.cpp/vulkan (speculative decode, dspark)), snapshot 2026-10-29.

Every measured single-stream run of DeepSeek-V4-Flash-0731 on Strix Halo

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
35.7	UD-IQ3_XXS	llama.cpp/vulkan (speculative decode, dspark)	524288	Structured table	pepuscz	2026-10-29	source
34.2	2.58bpw-mix	speculative decode, dspark		Extracted, source-linked	otheru	2026-08-29	source
29.6		rocm (speculative decode, dspark)	131072	Structured table	pepuscz	2026-10-29	source
19.8	UD-IQ3_XXS	llama.cpp/vulkan-radv-performance		Structured table	kyuz0	2026-10-29	source
16.2	UD-IQ2_XXS	llama.cpp/rocm-7.14		Structured table	kyuz0	2026-10-29	source
16.1	UD-IQ2_XXS	llama.cpp/rocm-7.2.4		Structured table	kyuz0	2026-10-29	source
16	UD-IQ3_XXS	llama.cpp/rocm-7.14		Structured table	kyuz0	2026-10-29	source
15.8	UD-IQ3_XXS	llama.cpp/rocm-7.2.4		Structured table	kyuz0	2026-10-29	source
14.8	UD-IQ2_XXS	llama.cpp/rocm-7.14-pr26592		Structured table	kyuz0	2026-09-01	source
14.5	UD-IQ3_XXS	llama.cpp/rocm-7.14-pr26592		Structured table	kyuz0	2026-09-01	source
13.1	UD-IQ2_XXS	llama.cpp/vulkan-radv-performance		Structured table	kyuz0	2026-10-29	source
12	UD-IQ2_XXS	vulkan	131072	Verified (ran it)	sypherin	2026-10-10	source
9.1	UD-IQ2_XXS	llama.cpp/vulkan-radv		Structured table	kyuz0	2026-10-29	source
9	UD-IQ3_XXS	llama.cpp/vulkan-radv		Structured table	kyuz0	2026-10-29	source

Best quant on Strix Halo

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

UD-IQ3_XXS: 19.8 tok/s, fastest (llama.cpp/vulkan-radv-performance)
UD-IQ2_XXS: 16.2 tok/s, 18% lower (llama.cpp/rocm-7.14)

Backends on Strix Halo

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

llama.cpp/vulkan-radv-performance: 19.8 tok/s (UD-IQ3_XXS)
llama.cpp/rocm-7.14: 16.2 tok/s (UD-IQ2_XXS)
llama.cpp/rocm-7.2.4: 16.1 tok/s (UD-IQ2_XXS)
llama.cpp/rocm-7.14-pr26592: 14.8 tok/s (UD-IQ2_XXS)
vulkan: 12 tok/s (UD-IQ2_XXS)
llama.cpp/vulkan-radv: 9.1 tok/s (UD-IQ2_XXS)

DeepSeek-V4-Flash-0731 on other boxes

DGX Spark: 32.8 tok/s best of 1 config

Questions and answers

How fast is DeepSeek-V4-Flash-0731 on Strix Halo?: The fastest of 14 measured configs is 35.7 tok/s single-stream (UD-IQ3_XXS, llama.cpp/vulkan (speculative decode, dspark)). The slowest is 9 tok/s.
Which quant of DeepSeek-V4-Flash-0731 is fastest on Strix Halo?: UD-IQ3_XXS at 19.8 tok/s, ahead of UD-IQ2_XXS at 16.2. Plain decoding only.
Which backend is fastest for DeepSeek-V4-Flash-0731 on Strix Halo?: llama.cpp/vulkan-radv-performance at 19.8 tok/s, then llama.cpp/rocm-7.14 at 16.2, llama.cpp/rocm-7.2.4 at 16.1, llama.cpp/rocm-7.14-pr26592 at 14.8, vulkan at 12, llama.cpp/vulkan-radv at 9.1. Each backend figure is its best plain-decoding row and may use a different quant.
How does Strix Halo compare with other boxes for DeepSeek-V4-Flash-0731?: Strix Halo best 35.7 tok/s. Other boxes, best measured: DGX Spark 32.8. Backends and quants differ, so this is the best known number per box, not a controlled test.

See also: DeepSeek-V4-Flash-0731 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.