How fast is Qwen3.5-9B on DGX Spark?

The fastest of 18 measured configs is 49.2 tok/s single-stream (UD-IQ3_XXS, llama.cpp/cuda). The slowest is 20 tok/s.

Which quant of Qwen3.5-9B is fastest on DGX Spark?

UD-IQ3_XXS at 49.2 tok/s, ahead of IQ4_XS at 42.4, Q3_K_S at 41.9, IQ4_NL at 41.1, Q4_0 at 41, Q3_K_M at 40.2, Q4_1 at 39.9, UD-Q3_K_XL at 39.2, Q4_K_S at 38.1, Q4_K_M at 36.4, UD-Q4_K_XL at 34.9, Q5_K_S at 32.7, Q5_K_M at 31.5, UD-Q5_K_XL at 31, Q6_K at 28.4, UD-Q6_K_XL at 26.2, Q8_0 at 25, UD-Q8_K_XL at 20. Plain decoding only.

Configs › Qwen3.5-9B › Qwen3.5-9B on DGX Spark: 49.2 tok/s, 18 measured configs

Qwen3.5-9B on DGX Spark: 49.2 tok/s, 18 measured configs

Measured Qwen3.5-9B speeds on DGX Spark: fastest 49.2 tok/s single-stream (UD-IQ3_XXS, llama.cpp/cuda), 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 Qwen3.5-9B on DGX Spark: 49.2 tok/s (UD-IQ3_XXS, llama.cpp/cuda), snapshot 2026-10-29.

Every measured single-stream run of Qwen3.5-9B on DGX Spark

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
49.2	UD-IQ3_XXS	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
42.4	IQ4_XS	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
41.9	Q3_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
41.1	IQ4_NL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
41	Q4_0	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
40.2	Q3_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
39.9	Q4_1	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
39.2	UD-Q3_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
38.1	Q4_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
36.4	Q4_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
34.9	UD-Q4_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
32.7	Q5_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
31.5	Q5_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
31	UD-Q5_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
28.4	Q6_K	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
26.2	UD-Q6_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
25	Q8_0	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
20	UD-Q8_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source

Best quant on DGX Spark

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

UD-IQ3_XXS: 49.2 tok/s, fastest (llama.cpp/cuda)
IQ4_XS: 42.4 tok/s, 14% lower (llama.cpp/cuda)
Q3_K_S: 41.9 tok/s, 15% lower (llama.cpp/cuda)
IQ4_NL: 41.1 tok/s, 16% lower (llama.cpp/cuda)
Q4_0: 41 tok/s, 17% lower (llama.cpp/cuda)
Q3_K_M: 40.2 tok/s, 18% lower (llama.cpp/cuda)
Q4_1: 39.9 tok/s, 19% lower (llama.cpp/cuda)
UD-Q3_K_XL: 39.2 tok/s, 20% lower (llama.cpp/cuda)
Q4_K_S: 38.1 tok/s, 23% lower (llama.cpp/cuda)
Q4_K_M: 36.4 tok/s, 26% lower (llama.cpp/cuda)
UD-Q4_K_XL: 34.9 tok/s, 29% lower (llama.cpp/cuda)
Q5_K_S: 32.7 tok/s, 34% lower (llama.cpp/cuda)
Q5_K_M: 31.5 tok/s, 36% lower (llama.cpp/cuda)
UD-Q5_K_XL: 31 tok/s, 37% lower (llama.cpp/cuda)
Q6_K: 28.4 tok/s, 42% lower (llama.cpp/cuda)
UD-Q6_K_XL: 26.2 tok/s, 47% lower (llama.cpp/cuda)
Q8_0: 25 tok/s, 49% lower (llama.cpp/cuda)
UD-Q8_K_XL: 20 tok/s, 59% lower (llama.cpp/cuda)

Questions and answers

How fast is Qwen3.5-9B on DGX Spark?: The fastest of 18 measured configs is 49.2 tok/s single-stream (UD-IQ3_XXS, llama.cpp/cuda). The slowest is 20 tok/s.
Which quant of Qwen3.5-9B is fastest on DGX Spark?: UD-IQ3_XXS at 49.2 tok/s, ahead of IQ4_XS at 42.4, Q3_K_S at 41.9, IQ4_NL at 41.1, Q4_0 at 41, Q3_K_M at 40.2, Q4_1 at 39.9, UD-Q3_K_XL at 39.2, Q4_K_S at 38.1, Q4_K_M at 36.4, UD-Q4_K_XL at 34.9, Q5_K_S at 32.7, Q5_K_M at 31.5, UD-Q5_K_XL at 31, Q6_K at 28.4, UD-Q6_K_XL at 26.2, Q8_0 at 25, UD-Q8_K_XL at 20. Plain decoding only.

See also: Qwen3.5-9B on every box · DGX Spark hardware page (from S$7,452 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.