Configs › Qwen3.5-27B on DGX Spark: 13.2 tok/s measured
Qwen3.5-27B on DGX Spark: 13.2 tok/s measured
Measured Qwen3.5-27B speeds: 13.2 tok/s on DGX Spark (IQ4_XS, 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 Qwen3.5-27B on DGX Spark?
- 13.2 tok/s single-stream decode (IQ4_XS, 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 Qwen3.5-27B on RTX 5090?
- 72.6 tok/s single-stream decode (UD-IQ3_XXS, llama.cpp/cuda), the fastest of 15 measured configs on NVIDIA RTX 5090 (32GB), measured by paulplee, source last checked 2026-10-15. source
- How fast is Qwen3.5-27B on H100 SXM?
- 68.5 tok/s single-stream decode (Q4_K_M, llama.cpp/cuda), the only measured config on NVIDIA H100 SXM (80GB), measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.5-27B on H100 PCIe?
- 52.3 tok/s single-stream decode (Q4_K_M, llama.cpp/cuda), the only measured config on NVIDIA H100 PCIe (80GB), measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.5-27B on A100 40GB?
- 42.4 tok/s single-stream decode (Q4_K_M, llama.cpp/cuda), the only measured config on NVIDIA A100 40GB, measured by sypherin, source last checked 2026-10-14. source
- How fast is Qwen3.5-27B on A10?
- 23.5 tok/s single-stream decode (Q4_K_M, llama.cpp/cuda), the only measured config on NVIDIA A10 (24GB), measured by sypherin, source last checked 2026-10-14. source
Measured configs for Qwen3.5-27B on AI boxes
| Model | Hardware | Quant | Backend | Decode | Context | Trust | Source |
|---|---|---|---|---|---|---|---|
| Qwen3.5-27B | DGX Spark | IQ4_XS | llama.cpp/cuda | 13.2 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q3_K_S | llama.cpp/cuda | 13 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | NVFP4 | atlas | 13 tok/s | Extracted, source-linked | source | |
| Qwen3.5-27B | DGX Spark | Q4_0 | llama.cpp/cuda | 12.7 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | IQ4_NL | llama.cpp/cuda | 12.7 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q3_K_M | llama.cpp/cuda | 12.6 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q4_1 | llama.cpp/cuda | 12.3 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | UD-Q3_K_XL | llama.cpp/cuda | 11.9 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q4_K_S | llama.cpp/cuda | 11.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q4_K_M | llama.cpp/cuda | 11.1 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | UD-Q4_K_XL | llama.cpp/cuda | 10.6 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q5_K_S | llama.cpp/cuda | 9.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q5_K_M | llama.cpp/cuda | 9.5 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | UD-Q5_K_XL | llama.cpp/cuda | 9.3 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q6_K | llama.cpp/cuda | 8.5 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | UD-Q6_K_XL | llama.cpp/cuda | 7.7 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | Q8_0 | llama.cpp/cuda | 7.4 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | DGX Spark | UD-Q8_K_XL | llama.cpp/cuda | 6.5 tok/s | 8192 | Structured table | source |
Qwen3.5-27B 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 |
|---|---|---|---|---|---|---|---|
| Qwen3.5-27B | RTX 5090 | UD-IQ3_XXS | llama.cpp/cuda | 72.6 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | H100 SXM | Q4_K_M | llama.cpp/cuda | 68.5 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.5-27B | RTX 5090 | IQ4_XS | llama.cpp/cuda | 67.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | IQ4_NL | llama.cpp/cuda | 66.2 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | Q4_K_S | llama.cpp/cuda | 65.5 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | Q4_0 | llama.cpp/cuda | 65.4 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | Q3_K_M | llama.cpp/cuda | 64.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | Q3_K_S | llama.cpp/cuda | 63.8 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | Q4_1 | llama.cpp/cuda | 63.3 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | Q4_K_M | llama.cpp/cuda | 62.6 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | UD-Q3_K_XL | llama.cpp/cuda | 62.2 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | UD-Q4_K_XL | llama.cpp/cuda | 59.7 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | Q5_K_S | llama.cpp/cuda | 58 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | Q5_K_M | llama.cpp/cuda | 56.6 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | RTX 5090 | UD-Q5_K_XL | llama.cpp/cuda | 55.2 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | H100 PCIe | Q4_K_M | llama.cpp/cuda | 52.3 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.5-27B | RTX 5090 | Q6_K | llama.cpp/cuda | 51.3 tok/s | 8192 | Structured table | source |
| Qwen3.5-27B | A100 40GB | Q4_K_M | llama.cpp/cuda | 42.4 tok/s | 8192 | Verified (ran it) | source |
| Qwen3.5-27B | A10 | Q4_K_M | llama.cpp/cuda | 23.5 tok/s | 8192 | Verified (ran it) | 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.