How fast is Qwen3.6-27B on DGX Spark?

The fastest of 9 measured configs is 37.3 tok/s single-stream (NVFP4, vllm (speculative decode, mtp-3)). The slowest is 9.9 tok/s.

Which quant of Qwen3.6-27B is fastest on DGX Spark?

Q3_K_S at 13.3 tok/s, ahead of IQ4_XS at 13.2, Q3_K_M at 12.8, Q4_K_S at 12.2, Q3_K_L at 12, Q4_K_M at 11.5, Q5_K_S at 10.2, Q5_K_M at 9.9. Plain decoding only.

How does DGX Spark compare with other boxes for Qwen3.6-27B?

DGX Spark best 37.3 tok/s. Other boxes, best measured: Strix Halo 21. Backends and quants differ, so this is the best known number per box, not a controlled test.

Configs › Qwen3.6-27B › Qwen3.6-27B on DGX Spark: 37.3 tok/s, 9 measured configs

Qwen3.6-27B on DGX Spark: 37.3 tok/s, 9 measured configs

Measured Qwen3.6-27B speeds on DGX Spark: fastest 37.3 tok/s single-stream (NVFP4, vllm (speculative decode, mtp-3)), 9 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 Qwen3.6-27B on DGX Spark: 37.3 tok/s (NVFP4, vllm (speculative decode, mtp-3)), snapshot 2026-10-29.

Every measured single-stream run of Qwen3.6-27B on DGX Spark

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
37.3	NVFP4	vllm (speculative decode, mtp-3)		Structured table	jvr0x	2026-09-22	source
13.3	Q3_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
13.2	IQ4_XS	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
12.8	Q3_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
12.2	Q4_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
12	Q3_K_L	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
11.5	Q4_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
10.2	Q5_K_S	llama.cpp/cuda	2048	Structured table	paulplee	2026-10-29	source
9.9	Q5_K_M	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.

Q3_K_S: 13.3 tok/s, fastest (llama.cpp/cuda)
IQ4_XS: 13.2 tok/s, 1% lower (llama.cpp/cuda)
Q3_K_M: 12.8 tok/s, 4% lower (llama.cpp/cuda)
Q4_K_S: 12.2 tok/s, 8% lower (llama.cpp/cuda)
Q3_K_L: 12 tok/s, 10% lower (llama.cpp/cuda)
Q4_K_M: 11.5 tok/s, 14% lower (llama.cpp/cuda)
Q5_K_S: 10.2 tok/s, 23% lower (llama.cpp/cuda)
Q5_K_M: 9.9 tok/s, 26% lower (llama.cpp/cuda)

Qwen3.6-27B on other boxes

Strix Halo: 21 tok/s best of 18 configs

Questions and answers

How fast is Qwen3.6-27B on DGX Spark?: The fastest of 9 measured configs is 37.3 tok/s single-stream (NVFP4, vllm (speculative decode, mtp-3)). The slowest is 9.9 tok/s.
Which quant of Qwen3.6-27B is fastest on DGX Spark?: Q3_K_S at 13.3 tok/s, ahead of IQ4_XS at 13.2, Q3_K_M at 12.8, Q4_K_S at 12.2, Q3_K_L at 12, Q4_K_M at 11.5, Q5_K_S at 10.2, Q5_K_M at 9.9. Plain decoding only.
How does DGX Spark compare with other boxes for Qwen3.6-27B?: DGX Spark best 37.3 tok/s. Other boxes, best measured: Strix Halo 21. Backends and quants differ, so this is the best known number per box, not a controlled test.

See also: Qwen3.6-27B 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.