How fast is Qwen3.6-35B-A3B on DGX Spark?

The fastest of 22 measured configs is 77.4 tok/s single-stream (NVFP4, vllm). The slowest is 43.4 tok/s.

Which quant of Qwen3.6-35B-A3B is fastest on DGX Spark?

NVFP4 at 77.4 tok/s, ahead of UD-IQ3_XXS at 66.3, UD-IQ3_S at 65.6, UD-Q3_K_XL at 64.1, UD-Q3_K_M at 64, UD-IQ4_NL at 63.8, UD-IQ4_XS at 63.3, UD-IQ4_NL_XL at 63.1, Q4_K_XL at 62.7, UD-Q4_K_S at 62.1, MXFP4 at 61.4, UD-Q4_K_M at 61, UD-Q3_K_S at 61, UD-Q4_K_XL at 59.7, UD-Q5_K_S at 57.8, UD-Q5_K_M at 57.5, UD-Q5_K_XL at 57.2, UD-Q6_K at 56, UD-Q6_K_XL at 51.1, Q8_0 at 50.2, UD-Q8_K_XL at 43.4. Plain decoding only.

Which backend is fastest for Qwen3.6-35B-A3B on DGX Spark?

vllm at 77.4 tok/s, then llama.cpp/cuda at 66.3, llama.cpp at 62.7. Each backend figure is its best plain-decoding row and may use a different quant.

How does DGX Spark compare with other boxes for Qwen3.6-35B-A3B?

DGX Spark best 77.4 tok/s. Other boxes, best measured: Mac Max 151.1, Strix Halo 127.1, Mac Ultra 109.1. Backends and quants differ, so this is the best known number per box, not a controlled test.

Configs › Qwen3.6-35B-A3B › Qwen3.6-35B-A3B on DGX Spark: 77.4 tok/s, 22 measured configs

Qwen3.6-35B-A3B on DGX Spark: 77.4 tok/s, 22 measured configs

Measured Qwen3.6-35B-A3B speeds on DGX Spark: fastest 77.4 tok/s single-stream (NVFP4, vllm), 22 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 Qwen3.6-35B-A3B on DGX Spark: 77.4 tok/s (NVFP4, vllm), snapshot 2026-10-29.

Every measured single-stream run of Qwen3.6-35B-A3B on DGX Spark

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
77.4	NVFP4	vllm		Structured table	jvr0x	2026-10-29	source
66.3	UD-IQ3_XXS	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
65.6	UD-IQ3_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
64.1	UD-Q3_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
64	UD-Q3_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
63.8	UD-IQ4_NL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
63.3	UD-IQ4_XS	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
63.1	UD-IQ4_NL_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
62.7	Q4_K_XL	llama.cpp		Structured table	jvr0x	2026-10-29	source
62.1	UD-Q4_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
61.4	MXFP4	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
61	UD-Q4_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
61	UD-Q3_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
59.7	UD-Q4_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
57.8	UD-Q5_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
57.5	UD-Q5_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
57.2	UD-Q5_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
56	UD-Q6_K	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
51.1	UD-Q6_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
50.2	Q8_0	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
48.2		llama.cpp		Structured table	jvr0x	2026-10-29	source
43.4	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.

NVFP4: 77.4 tok/s, fastest (vllm)
UD-IQ3_XXS: 66.3 tok/s, 14% lower (llama.cpp/cuda)
UD-IQ3_S: 65.6 tok/s, 15% lower (llama.cpp/cuda)
UD-Q3_K_XL: 64.1 tok/s, 17% lower (llama.cpp/cuda)
UD-Q3_K_M: 64 tok/s, 17% lower (llama.cpp/cuda)
UD-IQ4_NL: 63.8 tok/s, 18% lower (llama.cpp/cuda)
UD-IQ4_XS: 63.3 tok/s, 18% lower (llama.cpp/cuda)
UD-IQ4_NL_XL: 63.1 tok/s, 18% lower (llama.cpp/cuda)
Q4_K_XL: 62.7 tok/s, 19% lower (llama.cpp)
UD-Q4_K_S: 62.1 tok/s, 20% lower (llama.cpp/cuda)
MXFP4: 61.4 tok/s, 21% lower (llama.cpp/cuda)
UD-Q4_K_M: 61 tok/s, 21% lower (llama.cpp/cuda)
UD-Q3_K_S: 61 tok/s, 21% lower (llama.cpp/cuda)
UD-Q4_K_XL: 59.7 tok/s, 23% lower (llama.cpp/cuda)
UD-Q5_K_S: 57.8 tok/s, 25% lower (llama.cpp/cuda)
UD-Q5_K_M: 57.5 tok/s, 26% lower (llama.cpp/cuda)
UD-Q5_K_XL: 57.2 tok/s, 26% lower (llama.cpp/cuda)
UD-Q6_K: 56 tok/s, 28% lower (llama.cpp/cuda)
UD-Q6_K_XL: 51.1 tok/s, 34% lower (llama.cpp/cuda)
Q8_0: 50.2 tok/s, 35% lower (llama.cpp/cuda)
UD-Q8_K_XL: 43.4 tok/s, 44% lower (llama.cpp/cuda)

Backends on DGX Spark

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

vllm: 77.4 tok/s (NVFP4)
llama.cpp/cuda: 66.3 tok/s (UD-IQ3_XXS)
llama.cpp: 62.7 tok/s (Q4_K_XL)

Qwen3.6-35B-A3B on other boxes

Mac Max: 151.1 tok/s best of 3 configs
Strix Halo: 127.1 tok/s best of 29 configs
Mac Ultra: 109.1 tok/s best of 1 config

Questions and answers

How fast is Qwen3.6-35B-A3B on DGX Spark?: The fastest of 22 measured configs is 77.4 tok/s single-stream (NVFP4, vllm). The slowest is 43.4 tok/s.
Which quant of Qwen3.6-35B-A3B is fastest on DGX Spark?: NVFP4 at 77.4 tok/s, ahead of UD-IQ3_XXS at 66.3, UD-IQ3_S at 65.6, UD-Q3_K_XL at 64.1, UD-Q3_K_M at 64, UD-IQ4_NL at 63.8, UD-IQ4_XS at 63.3, UD-IQ4_NL_XL at 63.1, Q4_K_XL at 62.7, UD-Q4_K_S at 62.1, MXFP4 at 61.4, UD-Q4_K_M at 61, UD-Q3_K_S at 61, UD-Q4_K_XL at 59.7, UD-Q5_K_S at 57.8, UD-Q5_K_M at 57.5, UD-Q5_K_XL at 57.2, UD-Q6_K at 56, UD-Q6_K_XL at 51.1, Q8_0 at 50.2, UD-Q8_K_XL at 43.4. Plain decoding only.
Which backend is fastest for Qwen3.6-35B-A3B on DGX Spark?: vllm at 77.4 tok/s, then llama.cpp/cuda at 66.3, llama.cpp at 62.7. Each backend figure is its best plain-decoding row and may use a different quant.
How does DGX Spark compare with other boxes for Qwen3.6-35B-A3B?: DGX Spark best 77.4 tok/s. Other boxes, best measured: Mac Max 151.1, Strix Halo 127.1, Mac Ultra 109.1. Backends and quants differ, so this is the best known number per box, not a controlled test.

See also: Qwen3.6-35B-A3B 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.