← All benchmarks

Qwen3.6 35B-A3B

Paddock vs vLLM vs SGLang vs llama.cpp · Text Generation · RTX PRO 6000 Blackwell Max-Q (96 GB) · 2026-07-20

Hybrid Gated-DeltaNet + gated-attention MoE with a built-in MTP head for speculative decoding and a native vision encoder.

By AlibabaOpen model card ↗

Up to 1.6x

vs vLLM

Faster in 4 of 5 tests

Up to 2x

vs SGLang

Faster in 3 of 5 tests

Up to 4.4x

vs llama.cpp

Faster in 5 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. Every engine here uses speculative decoding: a small model guesses the next few words and the main model checks them in one go, so it produces several words at a time. Paddock does this by default, so the comparison is fair. These numbers rise and fall with the prompt and the load, and usually beat running without it. The dropdown shows the same engines with it off.

PaddockvLLMSGLangllama.cpp

Qwen3.6-35B-A3B

vs vLLM: 0.8-1.6xvs SGLang: 0.7-2xvs llama.cpp: 1.4-4.4x
1 client · 1k-token prompts · 256 tokens out
Paddock
292.3 tokens/s
vLLM
181.61.61x
SGLang
144.72.02x
llama.cpp
203.01.44x
4 clients · 256-token prompts · 1k tokens out
Paddock
636.8 tokens/s
vLLM
513.91.24x
SGLang
551.21.16x
llama.cpp
387.41.64x
8 clients · 1k-token prompts · 256 tokens out
Paddock
853.4 tokens/s
vLLM
767.71.11x
SGLang
826.31.03x
llama.cpp
325.42.62x
8 clients · 4k-token prompts · 128 tokens out
Paddock
383.2 tokens/s
vLLM
491.60.78x
SGLang
541.10.71x
llama.cpp
129.22.97x
32 clients · 1k-token prompts · 256 tokens out
Paddock
1,585.3 tokens/s
vLLM
1,568.91.01x
SGLang
1,602.50.99x
llama.cpp
356.54.45x
Show the same engines without speculative decoding

Speculative decoding is off by default in these engines - it needs a separate helper model and extra setup. This is how they run out of the box. Paddock always uses it, so its bar is unchanged and shown for scale.

Qwen3.6-35B-A3B

1 client · 1k-token prompts · 256 tokens out
Paddock
292.3 tokens/s
4 clients · 256-token prompts · 1k tokens out
Paddock
636.8 tokens/s
8 clients · 1k-token prompts · 256 tokens out
Paddock
853.4 tokens/s
8 clients · 4k-token prompts · 128 tokens out
Paddock
383.2 tokens/s
32 clients · 1k-token prompts · 256 tokens out
Paddock
1,585.3 tokens/s

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.

TestPaddockvLLMSGLangllama.cpp
1 client · 1k-token prompts · 256 tokens out83 ms88 ms1.1x85 ms1.0x365 ms4.4x
4 clients · 256-token prompts · 1k tokens out65 ms171 ms2.6x157 ms2.4x395 ms6.0x
8 clients · 1k-token prompts · 256 tokens out254 ms248 ms1.0x237 ms0.9x819 ms3.2x
8 clients · 4k-token prompts · 128 tokens out477 ms789 ms1.7x677 ms1.4x2,178 ms4.6x
32 clients · 1k-token prompts · 256 tokens out364 ms586 ms1.6x619 ms1.7x1,541 ms4.2x

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.

Qwen3.6-35B-A3B

PaddockvLLMSGLangllama.cpp
tokens/s05001,0001,5002,0001 client1k in · 256 out4 clients256 in · 1k out8 clients1k in · 256 out8 clients4k in · 128 out32 clients1k in · 256 outvLLM - 1 client · 1k-token prompts · 256 tokens out: 181.6 tokens/sSGLang - 1 client · 1k-token prompts · 256 tokens out: 144.7 tokens/sllama.cpp - 1 client · 1k-token prompts · 256 tokens out: 203.0 tokens/sPaddock - 1 client · 1k-token prompts · 256 tokens out: 292.3 tokens/s1.4×vLLM - 4 clients · 256-token prompts · 1k tokens out: 513.9 tokens/sSGLang - 4 clients · 256-token prompts · 1k tokens out: 551.2 tokens/sllama.cpp - 4 clients · 256-token prompts · 1k tokens out: 387.4 tokens/sPaddock - 4 clients · 256-token prompts · 1k tokens out: 636.8 tokens/s1.2×vLLM - 8 clients · 1k-token prompts · 256 tokens out: 767.7 tokens/sSGLang - 8 clients · 1k-token prompts · 256 tokens out: 826.3 tokens/sllama.cpp - 8 clients · 1k-token prompts · 256 tokens out: 325.4 tokens/sPaddock - 8 clients · 1k-token prompts · 256 tokens out: 853.4 tokens/s1.0×vLLM - 8 clients · 4k-token prompts · 128 tokens out: 491.6 tokens/sSGLang - 8 clients · 4k-token prompts · 128 tokens out: 541.1 tokens/sllama.cpp - 8 clients · 4k-token prompts · 128 tokens out: 129.2 tokens/sPaddock - 8 clients · 4k-token prompts · 128 tokens out: 383.2 tokens/s0.7×vLLM - 32 clients · 1k-token prompts · 256 tokens out: 1,568.9 tokens/sSGLang - 32 clients · 1k-token prompts · 256 tokens out: 1,602.5 tokens/sllama.cpp - 32 clients · 1k-token prompts · 256 tokens out: 356.5 tokens/sPaddock - 32 clients · 1k-token prompts · 256 tokens out: 1,585.3 tokens/s1.0×+89 tokens/s

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
0.25.1
SGLang
0.5.15.post1
llama.cpp (speculative)
b10069

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.