How fast is Qwen3.5-27B on DGX Spark?

The fastest of 18 measured configs is 13.2 tok/s single-stream (IQ4_XS, llama.cpp/cuda). The slowest is 6.5 tok/s.

Which quant of Qwen3.5-27B is fastest on DGX Spark?

IQ4_XS at 13.2 tok/s, ahead of NVFP4 at 13, Q3_K_S at 13, Q4_0 at 12.7, IQ4_NL at 12.7, Q3_K_M at 12.6, Q4_1 at 12.3, UD-Q3_K_XL at 11.9, Q4_K_S at 11.8, Q4_K_M at 11.1, UD-Q4_K_XL at 10.6, Q5_K_S at 9.8, Q5_K_M at 9.5, UD-Q5_K_XL at 9.3, Q6_K at 8.5, UD-Q6_K_XL at 7.7, Q8_0 at 7.4, UD-Q8_K_XL at 6.5. Plain decoding only.

Which backend is fastest for Qwen3.5-27B on DGX Spark?

llama.cpp/cuda at 13.2 tok/s, then atlas at 13. Each backend figure is its best plain-decoding row and may use a different quant.

Configs › Qwen3.5-27B › Qwen3.5-27B on DGX Spark: 13.2 tok/s, 18 measured configs

Qwen3.5-27B on DGX Spark: 13.2 tok/s, 18 measured configs

Measured Qwen3.5-27B speeds on DGX Spark: fastest 13.2 tok/s single-stream (IQ4_XS, 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 Qwen3.5-27B on DGX Spark: 13.2 tok/s (IQ4_XS, llama.cpp/cuda), snapshot 2026-10-29.

Every measured single-stream run of Qwen3.5-27B on DGX Spark

Decode tok/s	Quant	Backend	Context	Trust	Measured by	Source checked	Source
13.2	IQ4_XS	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
13	NVFP4	atlas		Extracted, source-linked	Atlas-Inf	2026-09-23	source
13	Q3_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
12.7	Q4_0	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
12.7	IQ4_NL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
12.6	Q3_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
12.3	Q4_1	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
11.9	UD-Q3_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
11.8	Q4_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
11.1	Q4_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
10.6	UD-Q4_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
9.8	Q5_K_S	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
9.5	Q5_K_M	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
9.3	UD-Q5_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
8.5	Q6_K	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
7.7	UD-Q6_K_XL	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
7.4	Q8_0	llama.cpp/cuda	8192	Structured table	paulplee	2026-10-29	source
6.5	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.

IQ4_XS: 13.2 tok/s, fastest (llama.cpp/cuda)
NVFP4: 13 tok/s, 2% lower (atlas)
Q3_K_S: 13 tok/s, 2% lower (llama.cpp/cuda)
Q4_0: 12.7 tok/s, 4% lower (llama.cpp/cuda)
IQ4_NL: 12.7 tok/s, 4% lower (llama.cpp/cuda)
Q3_K_M: 12.6 tok/s, 5% lower (llama.cpp/cuda)
Q4_1: 12.3 tok/s, 7% lower (llama.cpp/cuda)
UD-Q3_K_XL: 11.9 tok/s, 10% lower (llama.cpp/cuda)
Q4_K_S: 11.8 tok/s, 11% lower (llama.cpp/cuda)
Q4_K_M: 11.1 tok/s, 16% lower (llama.cpp/cuda)
UD-Q4_K_XL: 10.6 tok/s, 20% lower (llama.cpp/cuda)
Q5_K_S: 9.8 tok/s, 26% lower (llama.cpp/cuda)
Q5_K_M: 9.5 tok/s, 28% lower (llama.cpp/cuda)
UD-Q5_K_XL: 9.3 tok/s, 30% lower (llama.cpp/cuda)
Q6_K: 8.5 tok/s, 36% lower (llama.cpp/cuda)
UD-Q6_K_XL: 7.7 tok/s, 42% lower (llama.cpp/cuda)
Q8_0: 7.4 tok/s, 44% lower (llama.cpp/cuda)
UD-Q8_K_XL: 6.5 tok/s, 51% lower (llama.cpp/cuda)

Backends on DGX Spark

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

llama.cpp/cuda: 13.2 tok/s (IQ4_XS)
atlas: 13 tok/s (NVFP4)

Questions and answers

How fast is Qwen3.5-27B on DGX Spark?: The fastest of 18 measured configs is 13.2 tok/s single-stream (IQ4_XS, llama.cpp/cuda). The slowest is 6.5 tok/s.
Which quant of Qwen3.5-27B is fastest on DGX Spark?: IQ4_XS at 13.2 tok/s, ahead of NVFP4 at 13, Q3_K_S at 13, Q4_0 at 12.7, IQ4_NL at 12.7, Q3_K_M at 12.6, Q4_1 at 12.3, UD-Q3_K_XL at 11.9, Q4_K_S at 11.8, Q4_K_M at 11.1, UD-Q4_K_XL at 10.6, Q5_K_S at 9.8, Q5_K_M at 9.5, UD-Q5_K_XL at 9.3, Q6_K at 8.5, UD-Q6_K_XL at 7.7, Q8_0 at 7.4, UD-Q8_K_XL at 6.5. Plain decoding only.
Which backend is fastest for Qwen3.5-27B on DGX Spark?: llama.cpp/cuda at 13.2 tok/s, then atlas at 13. Each backend figure is its best plain-decoding row and may use a different quant.

See also: Qwen3.5-27B 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.