Qwen3.5 9B
Paddock vs vLLM vs SGLang vs llama.cpp · Text Generation · RTX 5090 (32 GB) · 2026-07-20
Qwen3.5-series model - a Gated DeltaNet + gated-attention hybrid focused on architectural efficiency and long context.
RTX 5090 · 2026-07-20
Up to 1.7x
vs vLLM
Faster in 5 of 6 tests
Up to 2x
vs SGLang
Faster in 5 of 6 tests
Up to 3.5x
vs llama.cpp
Faster in 6 of 6 tests
We measure throughput end to end and take the best timed round after a warmup. Each model is tested several ways, from one client sending large requests to many clients hitting the server at once. Every input is unique, so caching can't inflate anyone's numbers.
Measured Throughput
Each engine ran the same tests, in tokens/s. Longer bars are better; the number next to a competitor is how many times faster Paddock was.
Qwen3.5-9B
Time to First Token
How long a client waits for the first token (median, best round), in milliseconds. Lower is better. The multiplier is how many times lower Paddock's latency is than that engine's on the same test.
| Test | Paddock | vLLM | SGLang | llama.cpp |
|---|---|---|---|---|
| 1 client · 1k-token prompts · 256 tokens out | 75 ms | 95 ms1.3x | 75 ms1.0x | 191 ms2.6x |
| 4 clients · 256-token prompts · 1k tokens out | 74 ms | 91 ms1.2x | 184 ms2.5x | 485 ms6.6x |
| 8 clients · 1k-token prompts · 256 tokens out | 282 ms | 384 ms1.4x | 410 ms1.5x | 1,474 ms5.2x |
| 8 clients · 4k-token prompts · 128 tokens out | 516 ms | 611 ms1.2x | 1,400 ms2.7x | 1,802 ms3.5x |
| 16 clients · agentic-token prompts · 256 tokens out | 529 ms | 356 ms0.7x | 467 ms0.9x | 2,509 ms4.7x |
| 32 clients · 1k-token prompts · 256 tokens out | 404 ms | 495 ms1.2x | 4,671 ms11.6x | 1,901 ms4.7x |
Throughput Under Load
The same results drawn as lines, from a single client on the left to the most concurrent test on the right. The shaded area is the gap between Paddock and the strongest competitor at each point, and the small numbers are Paddock's speedup there.
Qwen3.5-9B
Test Environment
Hardware
- GPU
- RTX 5090, 32 GB
Engine Versions
- Paddock
- pre-release build
- vLLM
- 0.25.1
- SGLang
- 0.5.15.post1
- llama.cpp
- latest master @ 178a6c44 (~b10069)
Each engine ran the best setup it supports on this hardware, with everything on the GPU.
RTX PRO 6000 Blackwell Max-Q · 2026-07-17
Up to 1.7x
vs vLLM
Faster in 4 of 5 tests
Up to 1.7x
vs SGLang
Faster in 4 of 5 tests
We measure throughput end to end and take the best timed round after a warmup. Each model is tested several ways, from one client sending large requests to many clients hitting the server at once. Every input is unique, so caching can't inflate anyone's numbers.
Measured Throughput
Each engine ran the same tests, in tokens/s. Longer bars are better; the number next to a competitor is how many times faster Paddock was.
Qwen3.5-9B
Time to First Token
How long a client waits for the first token (median, best round), in milliseconds. Lower is better. The multiplier is how many times lower Paddock's latency is than that engine's on the same test.
| Test | Paddock | vLLM | SGLang |
|---|---|---|---|
| 1 client · 1k-token prompts · 256 tokens out | 98 ms | 100 ms1.0x | 77 ms0.8x |
| 4 clients · 256-token prompts · 1k tokens out | 94 ms | 103 ms1.1x | 92 ms1.0x |
| 8 clients · 1k-token prompts · 256 tokens out | 360 ms | 448 ms1.2x | 389 ms1.1x |
| 8 clients · 4k-token prompts · 128 tokens out | 681 ms | 1,725 ms2.5x | 1,232 ms1.8x |
| 32 clients · 1k-token prompts · 256 tokens out | 521 ms | 1,377 ms2.6x | 1,331 ms2.6x |
Throughput Under Load
The same results drawn as lines, from a single client on the left to the most concurrent test on the right. The shaded area is the gap between Paddock and the strongest competitor at each point, and the small numbers are Paddock's speedup there.
Qwen3.5-9B
Test Environment
Hardware
- GPU
- RTX PRO 6000 Blackwell Max-Q, 96 GB
- Driver
- 595.71.05
- CUDA
- 13.0
- OS
- Ubuntu 24.04, Linux 6.8.0-117
Engine Versions
- Paddock
- pre-release build
- vLLM
- 0.25.0 (torch cu130)
- SGLang
- 0.5.15 (torch cu130)
Each engine ran the best setup it supports on this hardware, with everything on the GPU.
How We Tested
- Every engine ran on the same machine and GPU, never at the same time.
- The same test program sent identical requests to each engine.
- Every input was unique, so nothing could be served from a cache.
- Each number is the best timed round, after a warmup.
- We measure end to end: what an application connecting over HTTP actually gets.
- Where the engines differ in ways that could affect the comparison, we note it alongside the results.
Check our numbers
Paddock ships a benchmark harness, so you can run the same comparison on your own hardware.