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

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

Measured Qwen3.5-9B speeds: 49.2 tok/s on DGX Spark (UD-IQ3_XXS, llama.cpp/cuda). 18 configs, each linked to its source, snapshot 2026-10-15.

Snapshot 2026-10-15 · 394 configs on AI boxes and 423 on GPU cards · 123 models. Every number is generated from the tracker snapshot and linked to where it was measured; nothing here is typed by hand.

Quick answers

How fast is Qwen3.5-9B on DGX Spark?: 49.2 tok/s single-stream decode (UD-IQ3_XXS, llama.cpp/cuda), the fastest of 18 measured configs on NVIDIA DGX Spark (GB10), measured by paulplee, source last checked 2026-10-15. source
How fast is Qwen3.5-9B on RTX 5090?: 187 tok/s single-stream decode (UD-IQ3_XXS, llama.cpp/cuda), the fastest of 19 measured configs on NVIDIA RTX 5090 (32GB), measured by paulplee, source last checked 2026-10-15. source

Measured configs for Qwen3.5-9B on AI boxes

Model	Hardware	Quant	Backend	Decode	Context	Trust	Source
Qwen3.5-9B	DGX Spark	UD-IQ3_XXS	llama.cpp/cuda	49.2 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	IQ4_XS	llama.cpp/cuda	42.4 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q3_K_S	llama.cpp/cuda	41.9 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	IQ4_NL	llama.cpp/cuda	41.1 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q4_0	llama.cpp/cuda	41 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q3_K_M	llama.cpp/cuda	40.2 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q4_1	llama.cpp/cuda	39.9 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	UD-Q3_K_XL	llama.cpp/cuda	39.2 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q4_K_S	llama.cpp/cuda	38.1 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q4_K_M	llama.cpp/cuda	36.4 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	UD-Q4_K_XL	llama.cpp/cuda	34.9 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q5_K_S	llama.cpp/cuda	32.7 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q5_K_M	llama.cpp/cuda	31.5 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	UD-Q5_K_XL	llama.cpp/cuda	31 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q6_K	llama.cpp/cuda	28.4 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	UD-Q6_K_XL	llama.cpp/cuda	26.2 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	Q8_0	llama.cpp/cuda	25 tok/s	8192	Structured table	source
Qwen3.5-9B	DGX Spark	UD-Q8_K_XL	llama.cpp/cuda	20 tok/s	8192	Structured table	source

Qwen3.5-9B on GPU cards, ranked on their own

GPU cards run the model from their own VRAM: very fast on what fits, but most cannot load what a 128 GB AI box can. Ranked on their own, never against the boxes.

Model	Hardware	Quant	Backend	Decode	Context	Trust	Source
Qwen3.5-9B	RTX 5090	UD-IQ3_XXS	llama.cpp/cuda	187 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	IQ4_XS	llama.cpp/cuda	173.9 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q4_0	llama.cpp/cuda	169 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	IQ4_NL	llama.cpp/cuda	168.8 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q4_K_S	llama.cpp/cuda	167.3 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q4_1	llama.cpp/cuda	163.7 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	UD-Q3_K_XL	llama.cpp/cuda	162.8 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q4_K_M	llama.cpp/cuda	160.9 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q3_K_M	llama.cpp/cuda	160.5 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q3_K_S	llama.cpp/cuda	158.4 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	UD-Q4_K_XL	llama.cpp/cuda	157.3 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q5_K_S	llama.cpp/cuda	154.7 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q5_K_M	llama.cpp/cuda	151.1 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	UD-Q5_K_XL	llama.cpp/cuda	149.2 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q6_K	llama.cpp/cuda	139.7 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	UD-Q6_K_XL	llama.cpp/cuda	129.7 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	Q8_0	llama.cpp/cuda	123.9 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	UD-Q8_K_XL	llama.cpp/cuda	105.3 tok/s	8192	Structured table	source
Qwen3.5-9B	RTX 5090	BF16	llama.cpp/cuda	80.6 tok/s	8192	Structured table	source

More: all configs · hardware · news · contributors · methodology · llms.txt · llms-full.txt · API (OpenAPI) · JSON snapshot. Built by Altronis, private on-prem AI, Singapore.