← All benchmarks

Qwen3.6 27B

Paddock vs vLLM vs SGLang vs llama.cppRTX PRO 60002026-08-08

Dense 27B multimodal model tuned for flagship-level agentic coding, with a built-in MTP head for speculative decoding.

By AlibabaOpen model card ↗

Result

Throughput across the 13 fixed test scenarios. Paddock, vLLM and SGLang all serve the byte-identical official FP8 checkpoint. llama.cpp has no FP8 weight type to serve it with, so an FP8 checkpoint has to be dequantised when it is converted (llama.cpp's own converter has a --fp8-as-q8 flag for exactly that). It is measured here against the Paddock row reading the same Q8_0 file it reads. Speculative decoding is off.

vs vLLM

11 of 13

scenarios faster, up to +9.0% - behind on 2 · FP8 weights

vs SGLang

9 of 13

scenarios faster, up to +7.3% - behind on 2, level on 2 · FP8 weights

vs llama.cpp

13 of 13

scenarios faster, 1.16x-17.1x · Q8_0 weights

Throughput as the Load Climbs

What the server delivers, from one client up to a full batch. Higher is better. The shaded region is what Paddock delivers above the strongest of the others at each step, and the figure above the line is its multiple there.

Balanced load

1,024 tokens in · 1,024 tokens out

PaddockvLLMSGLangllama.cppQ8_0

Server throughput (tokens/s)

BESTWORST

The 64-client step is left off: past a max batch of 32 every engine is queueing rather than serving. It is in the table.

Table view
ClientsPaddock (FP8)vLLM (FP8)SGLang (FP8)llama.cpp (Q8_0)
147.7 tok/s45.0 tok/s45.0 tok/s33.7 tok/s
293.9 tok/s83.9 tok/s85.1 tok/s48.0 tok/s
4172.6 tok/s162.5 tok/s167.1 tok/s60.4 tok/s
8334.1 tok/s321.1 tok/s321.9 tok/s64.9 tok/s
16591.1 tok/s578.7 tok/s581.0 tok/s109.7 tok/s
32968.1 tok/s968.1 tok/s835.0 tok/s239.1 tok/s
64970.4 tok/s963.9 tok/s899.3 tok/s265.2 tok/s

Agentic sessions

Shared-prefix sessions × 6 turns

PaddockvLLMSGLangllama.cppQ8_0

Server throughput (tokens/s)

BESTWORST

Sessions share a prefix, so this ladder also measures whether an engine’s prefix cache engages.

Table view
SessionsPaddock (FP8)vLLM (FP8)SGLang (FP8)llama.cpp (Q8_0)
147.3 tok/s42.5 tok/s44.6 tok/s16.2 tok/s
293.1 tok/s75.3 tok/s83.5 tok/s19.2 tok/s
4171.9 tok/s135.8 tok/s164.0 tok/s22.2 tok/s
8333.2 tok/s233.3 tok/s309.7 tok/s24.2 tok/s
16598.9 tok/s363.5 tok/s562.7 tok/s32.6 tok/s
32970.3 tok/s492.5 tok/s921.6 tok/s120.2 tok/s

Total Throughput Against Per-User Speed

The trade every serving engine makes: pack in more clients and the server produces more tokens in total, but each client’s answer streams out slower. This is the curve the independent inference benchmarks publish - per-user speed across the bottom, total throughput up the side. Both axes are better when higher, so up and to the right wins. An engine is only genuinely better if it serves more tokens overall and streams faster to each client.

Balanced load

1,024 tokens in · 1,024 tokens out · each point is a client count

PaddockvLLMSGLangllama.cppQ8_0

Server throughput (tokens/s)

BESTWORST

Numbers on the Paddock line are client counts. llama.cpp is dashed because it serves Q8_0 rather than FP8.

Table view
ClientsPaddock (FP8)vLLM (FP8)SGLang (FP8)llama.cpp (Q8_0)
147.7 tok/s · 47.9/client45.0 tok/s · 45.2/client45.0 tok/s · 44.9/client33.7 tok/s · 45.5/client
293.9 tok/s · 47.4/client83.9 tok/s · 42.4/client85.1 tok/s · 43.1/client48.0 tok/s · 42.1/client
4172.6 tok/s · 43.9/client162.5 tok/s · 41.3/client167.1 tok/s · 42.2/client60.4 tok/s · 37.6/client
8334.1 tok/s · 42.7/client321.1 tok/s · 41.5/client321.9 tok/s · 41.2/client64.9 tok/s · 17.5/client
16591.1 tok/s · 37.8/client578.7 tok/s · 38.1/client581.0 tok/s · 37.7/client109.7 tok/s · 9.7/client
32968.1 tok/s · 31.0/client968.1 tok/s · 32.7/client835.0 tok/s · 32.8/client239.1 tok/s · 10.4/client
64970.4 tok/s · 30.8/client963.9 tok/s · 31.9/client899.3 tok/s · 32.6/client265.2 tok/s · 9.9/client

Waiting for the First Token

How long a client waits before anything appears, as the load climbs. The line is the median wait; the 90th percentile is in the table beneath it. Lower is better, and the axis is logarithmic - each gridline is ten times the one below it, because no straight scale holds a tenth of a second and ten seconds in one frame.

Balanced load

1,024 tokens in · 1,024 tokens out

PaddockvLLMSGLangllama.cppQ8_0

Time to first token, log scale

BESTWORST
Table view
ClientsPaddock (FP8)vLLM (FP8)SGLang (FP8)llama.cpp (Q8_0)
1120 ms · p90 120 ms146 ms · p90 149 ms114 ms · p90 115 ms8.0 s · p90 8.3 s
2240 ms · p90 244 ms232 ms · p90 278 ms190 ms · p90 200 ms19 s · p90 19 s
4403 ms · p90 425 ms367 ms · p90 522 ms311 ms · p90 353 ms41 s · p90 41 s
8549 ms · p90 727 ms694 ms · p90 996 ms656 ms · p90 683 ms74 s · p90 75 s
16635 ms · p90 1.3 s1.5 s · p90 1.6 s1.1 s · p90 1.3 s24 s · p90 132 s
32697 ms · p90 2.3 s2.5 s · p90 3.1 s1.9 s · p90 2.5 s6.9 s · p90 146 s
6434 s · p90 35 s36 s · p90 37 s35 s · p90 36 s115 s · p90 239 s

Agentic sessions

Shared-prefix sessions × 6 turns

PaddockvLLMSGLangllama.cppQ8_0

Time to first token, log scale

BESTWORST
Table view
SessionsPaddock (FP8)vLLM (FP8)SGLang (FP8)llama.cpp (Q8_0)
182 ms · p90 83 ms359 ms · p90 549 ms86 ms · p90 97 ms10 s · p90 10 s
2100 ms · p90 119 ms670 ms · p90 879 ms143 ms · p90 146 ms20 s · p90 21 s
4106 ms · p90 156 ms1.2 s · p90 1.6 s176 ms · p90 208 ms38 s · p90 41 s
8145 ms · p90 210 ms1.9 s · p90 2.3 s266 ms · p90 324 ms36 s · p90 78 s
16165 ms · p90 242 ms2.4 s · p90 3.2 s408 ms · p90 490 ms43 s · p90 138 s
32245 ms · p90 351 ms2.7 s · p90 3.9 s646 ms · p90 1.0 s5.4 s · p90 162 s

Agentic sessions share a prefix on purpose - reuse is the thing being measured. Paddock reports a cache hit rate of 0.95-0.98 and llama.cpp 0.88-0.97. vLLM’s prefix cache reports 0.0 here, and the wait shows it - vLLM climbs to 2.7 s at the top of the ladder. SGLang does not report cache statistics by default, so its 0.0 is unreported rather than disproven.

Every Shape on the Board

The four prompt/answer shapes NVIDIA, vLLM and SGLang publish, at 1, 8 and 32 concurrent clients, plus a vision probe. Longer bars are better, and each panel is scaled to itself. The badge on an engine is how many times faster Paddock is on that scenario - a grey badge below 1.0x means that engine won it. Each is rated against the Paddock row in its own weight class: llama.cpp against the row reading the identical Q8_0 file, vLLM and SGLang against the FP8 row.

Paddock (FP8)Paddock (Q8_0)vLLM (FP8)SGLang (FP8)llama.cpp (Q8_0)

Short chat

128 tokens in · 128 tokens outShort turns, where per-request overhead shows up.

Summarization

2,048 tokens in · 128 tokens outPrefill-bound: most of the work is reading the prompt.

Long generation

128 tokens in · 2,048 tokens outDecode-bound: most of the work is writing the answer.

Balanced

1,024 tokens in · 1,024 tokens outThe even mix, halfway between reading the prompt and writing the answer.

Vision document Q&A

One page image + questionA controlled probe of the vision path, not an industry dataset.

Where Paddock Wins and Where It Loses

The same scenarios as a margin against Paddock. The centre line is Paddock; a bar to the right means Paddock is ahead, a bar to the left means it is behind. Losses are labelled - they are the reason to read this chart.

Against vLLM and SGLang

Throughput per scenario, relative to Paddock on the byte-identical FP8 checkpoint

vLLMSGLang
WORSTBEST

The scale snaps to a fixed ladder rather than fitting the data - ±20% here - so a small margin looks small. Every cell Paddock drops is at 32 clients at once, against vLLM and SGLang.

Against llama.cpp

Same Q8_0 GGUF file, both engines - how many times faster Paddock is

Requests That Actually Met the Target

Throughput counts tokens whether or not anyone is still waiting for them. Goodput counts only the requests that met both halves of a declared target - first token within 2.0 s and 50 ms per token after that. It is the number that decides how many clients a server can really take. Higher is better.

Goodput under load

1,024 tokens in · 1,024 tokens out, target 2.0 s / 50 ms

PaddockvLLMSGLangllama.cppQ8_0

Requests meeting the SLO

BESTWORST
Table view
ClientsPaddock (FP8)vLLM (FP8)SGLang (FP8)llama.cpp (Q8_0)
1100.0%100.0%100.0%0.0%
2100.0%100.0%100.0%0.0%
4100.0%100.0%100.0%0.0%
8100.0%100.0%100.0%0.0%
16100.0%100.0%100.0%0.0%
3287.9%25.0%52.3%0.0%
647.0%4.3%8.2%0.0%

One scenario costs us this target: summarization · 32 clients, where the admission wave blows the 2.0 s budget (Paddock 8.3%, vLLM 0.5%, SGLang 2.6%, llama.cpp 0.0%). Under the goodput lens the ranking can flip against raw throughput. We publish it because it points at a real thing to fix.

The Whole Board

Every scenario, every engine, one table - throughput with median time to first token beneath it.

ScenarioPaddock (FP8)Paddock (Q8_0)vLLM (FP8)SGLang (FP8)llama.cpp (Q8_0)
Short chat · 1 client47.3 · 61 ms47.2 · 61 ms43.9 · 98 ms44.1 · 76 ms40.7 · 348 ms
Short chat · 8 clients325.6 · 150 ms326.1 · 149 ms314.1 · 229 ms323.1 · 122 ms148.1 · 2.8 s
Short chat · 32 clients942.2 · 224 ms944.2 · 223 ms967.8 · 453 ms996.6 · 395 ms155.6 · 9.8 s
Summarization · 1 client44.9 · 196 ms45.0 · 195 ms42.3 · 267 ms42.6 · 190 ms11.9 · 8.0 s
Summarization · 8 clients232.3 · 659 ms232.5 · 659 ms213.2 · 1.5 s231.6 · 1.1 s13.6 · 34 s
Summarization · 32 clients424.2 · 730 ms425.1 · 729 ms417.5 · 2.4 s426.0 · 3.6 s64.8 · 4.5 s
Long generation · 1 client47.9 · 61 ms47.9 · 61 ms45.2 · 101 ms44.9 · 82 ms37.0 · 10 s
Long generation · 8 clients343.5 · 150 ms343.4 · 148 ms331.0 · 219 ms329.6 · 122 ms119.1 · 70 s
Long generation · 32 clients1,048.7 · 241 ms1,048.8 · 246 ms1,068.5 · 465 ms1,058.6 · 435 ms344.3 · 11 s
Balanced · 1 client47.7 · 120 ms47.7 · 120 ms45.0 · 146 ms44.9 · 114 ms32.9 · 8.7 s
Balanced · 8 clients334.3 · 530 ms334.2 · 528 ms319.9 · 956 ms322.5 · 653 ms71.4 · 73 s
Vision document Q&A · 1 client45.2 · 185 ms45.2 · 184 ms43.4 · 143 ms43.8 · 92 ms11.1 · 8.9 s
Vision document Q&A · 8 clients307.7 · 390 ms307.7 · 389 ms285.6 · 578 ms304.5 · 295 ms21.7 · 44 s

How It Was Run

Machine and class

GPU
RTX PRO 6000 (sm_120, 96 GB, 600 W)
Date
2026-08-08
Class
speculative decoding off on every engine, read back from each serve config, 8-bit KV everywhere (paddock/vLLM/SGLang fp8-e4m3; llama.cpp q8_0 - ggml has no fp8 KV), equal-length generations, closed-loop, one engine at a time, same client
Target
First token ≤ 2.0 s, 50 ms per token after

One engine on the GPU at a time, verified idle in between, ~40 s warmup before the first measurement, same client driving all 5. Speculative decoding is off on every engine, including Paddock, so nothing is compared against an engine running without it.

Memory reserved

Paddock (FP8)
74.8 GB peak
Paddock (Q8_0)
74.7 GB peak
vLLM (FP8)
82.7 GB peak
SGLang (FP8)
91.1 GB peak
llama.cpp (Q8_0)
41.4 GB peak

Reserved, not consumed: each engine takes the share of the card its own memory setting tells it to.

Configuration

Paddock (FP8)
7318490e2026-08-07FP8 byte-passthrough (byte-identical checkpoint to vLLM/SGLang) + fp8-e4m3 KV + EXACT f32 GDN state, --no-spec --max-batch 32 --max-ctx 8192
Paddock (Q8_0)
7318490e2026-08-07Q8_0 GGUF default stack (same GGUF file llama.cpp serves) + fp8-e4m3 KV + EXACT f32 GDN state, --no-spec --max-batch 32 --max-ctx 8192
vLLM (FP8)
v0.26.0newest releaseofficial FP8 safetensors, --kv-cache-dtype fp8, --max-model-len 8192 --gpu-memory-utilization 0.85 --max-num-seqs 32 (speculative decoding off per its own log)
SGLang (FP8)
main@0ba46c88e2026-08-03; newer than release v0.5.16official FP8 safetensors, --kv-cache-dtype fp8_e4m3, --context-length 8192 --max-running-requests 32 --mem-fraction-static 0.85 (speculative decoding off per its own log)
llama.cpp (Q8_0)
b10307fc3f10b38, newest release tagsame Q8_0 GGUF as the Paddock Q8_0 row + mmproj-BF16, -ngl 99 -fa on -c 262144 -np 32 -ctk q8_0 -ctv q8_0 --jinja

Check our numbers

Paddock is a plain OpenAI endpoint, so you can put these numbers to NVIDIA's aiperf on your own hardware, with a tool that is nobody's home advantage.