← All benchmarks

Gemma 4 31B

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

Google's strongest Gemma 4 model - dense multimodal transformer with 5:1 sliding-window/global attention and a separate MTP drafter for speculative decoding.

By GoogleOpen model card ↗

Up to 1.8x

vs vLLM

Faster in 5 of 5 tests

Up to 2.2x

vs SGLang

Faster in 5 of 5 tests

Up to 9.4x

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

gemma-4-31B-it

vs vLLM: 1-1.8xvs SGLang: 1.5-2.2xvs llama.cpp: 1.1-9.4x
1 client · 1k-token prompts · 256 tokens out
Paddock
87.2 tokens/s
vLLM
49.11.78x
SGLang
40.12.17x
llama.cpp
52.01.68x
4 clients · 256-token prompts · 1k tokens out
Paddock
253.0 tokens/s
vLLM
183.41.38x
SGLang
153.21.65x
llama.cpp
222.81.14x
8 clients · 1k-token prompts · 256 tokens out
Paddock
294.9 tokens/s
vLLM
282.01.05x
SGLang
187.01.58x
llama.cpp
86.43.41x
8 clients · 4k-token prompts · 128 tokens out
Paddock
100.7 tokens/s
vLLM
94.51.07x
SGLang
66.11.52x
llama.cpp
23.94.21x
32 clients · 1k-token prompts · 256 tokens out
Paddock
464.8 tokens/s
vLLM
422.01.1x
SGLang
259.91.79x
llama.cpp
49.29.45x
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.

gemma-4-31B-it

1 client · 1k-token prompts · 256 tokens out
Paddock
87.2 tokens/s
vLLM
21.74.02x
SGLang
21.54.06x
llama.cpp
31.92.73x
4 clients · 256-token prompts · 1k tokens out
Paddock
253.0 tokens/s
vLLM
84.13.01x
SGLang
88.12.87x
llama.cpp
117.92.15x
8 clients · 1k-token prompts · 256 tokens out
Paddock
294.9 tokens/s
vLLM
138.62.13x
SGLang
143.32.06x
llama.cpp
54.45.42x
8 clients · 4k-token prompts · 128 tokens out
Paddock
100.7 tokens/s
vLLM
70.41.43x
SGLang
61.71.63x
llama.cpp
22.14.56x
32 clients · 1k-token prompts · 256 tokens out
Paddock
464.8 tokens/s
vLLM
272.21.71x
SGLang
274.21.7x
llama.cpp
44.910.35x

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 out289 ms309 ms1.1x329 ms1.1x2,021 ms7.0x
4 clients · 256-token prompts · 1k tokens out233 ms196 ms0.8x224 ms1.0x1,797 ms7.7x
8 clients · 1k-token prompts · 256 tokens out597 ms379 ms0.6x455 ms0.8x4,305 ms7.2x
8 clients · 4k-token prompts · 128 tokens out2,009 ms2,074 ms1.0x4,943 ms2.5x12,129 ms6.0x
32 clients · 1k-token prompts · 256 tokens out1,451 ms4,357 ms3.0x5,788 ms4.0x12,100 ms8.3x

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.

gemma-4-31B-it

PaddockvLLMSGLangllama.cpp
tokens/s01252503755001 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: 49.1 tokens/sSGLang - 1 client · 1k-token prompts · 256 tokens out: 40.1 tokens/sllama.cpp - 1 client · 1k-token prompts · 256 tokens out: 52.0 tokens/sPaddock - 1 client · 1k-token prompts · 256 tokens out: 87.2 tokens/s1.7×vLLM - 4 clients · 256-token prompts · 1k tokens out: 183.4 tokens/sSGLang - 4 clients · 256-token prompts · 1k tokens out: 153.2 tokens/sllama.cpp - 4 clients · 256-token prompts · 1k tokens out: 222.8 tokens/sPaddock - 4 clients · 256-token prompts · 1k tokens out: 253.0 tokens/s1.1×vLLM - 8 clients · 1k-token prompts · 256 tokens out: 282.0 tokens/sSGLang - 8 clients · 1k-token prompts · 256 tokens out: 187.0 tokens/sllama.cpp - 8 clients · 1k-token prompts · 256 tokens out: 86.4 tokens/sPaddock - 8 clients · 1k-token prompts · 256 tokens out: 294.9 tokens/s1.0×vLLM - 8 clients · 4k-token prompts · 128 tokens out: 94.5 tokens/sSGLang - 8 clients · 4k-token prompts · 128 tokens out: 66.1 tokens/sllama.cpp - 8 clients · 4k-token prompts · 128 tokens out: 23.9 tokens/sPaddock - 8 clients · 4k-token prompts · 128 tokens out: 100.7 tokens/s1.1×vLLM - 32 clients · 1k-token prompts · 256 tokens out: 422.0 tokens/sSGLang - 32 clients · 1k-token prompts · 256 tokens out: 259.9 tokens/sllama.cpp - 32 clients · 1k-token prompts · 256 tokens out: 49.2 tokens/sPaddock - 32 clients · 1k-token prompts · 256 tokens out: 464.8 tokens/s1.1×+43 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
vLLM (speculative)
0.25.1
SGLang
0.5.15.post1
SGLang (speculative)
0.5.15.post1
llama.cpp
b10068
llama.cpp (speculative)
b10068

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.