शुरुआत करेंलीडरबोर्डDecode calculatorमॉडलOrganizationsReportsMatrix runsहार्डवेयरबेंचमार्कमार्केटप्लेसरेंटलProAPI दस्तावेज़
भाषा
Actual Computer — Every computer, one endpoint

Speed test matrix

DeepSeek-V4.1-Flash-1017 on RTX PRO 6000 Blackwell Server Edition 384GB ×4

vllm · 0.32.0 · MXFP4+FP88 contexts (8K–1019K) × 8 concurrency levels (1–128) · prefill at 8K, 16K, 32K, 64K, 128K, 256K, 512K, 1019K, 1,048,512
by @alvarobarttlmx v0.1.60; llm-inference-bench 52216d7 + prefill-window patch (only lifts 128k prefill cap)Measured

Hardware

Type
Discrete GPU
GPU
RTX PRO 6000 Blackwell Server Edition
GPU count
4
VRAM
384 GB (4 × 96 GB)
CPU
Intel(R) Xeon(R) Platinum 8559C
RAM
2000 GB
OS
Ubuntu 24.04.3 LTS

Engine

Engine
vllm
Version
0.32.0
Quantization
MXFP4+FP8
Backend
cuda
Build
ghcr.io/alvarobartt/vllm@sha256:ebdd9351069705b01e49d3fe080a427729c5680c5fd87d13c7fef163d041838a
Commit
889b538fc509

Run

Served model
deepseek-ai/DeepSeek-V4.1-Flash-1017
Revision
1bae0e1088993b9b82de08f59de5be5a34e42170
Decode test
Fixed duration
Prefill mode
Separate cold-cache pass
Time per measurement
30 s
Max output
4,096 tokens

Launch command

vllm serve /root/.cache/huggingface/hub/models--deepseek-ai--DeepSeek-V4.1-Flash-1017/snapshots/1bae0e1088993b9b82de08f59de5be5a34e42170 --revision 1bae0e1088993b9b82de08f59de5be5a34e42170 --tokenizer-mode deepseek_v41 --tensor-parallel-size 4 --tool-call-parser deepseek_v41 --enable-auto-tool-choice --reasoning-parser deepseek_v41 --mm-encoder-tp-mode data --block-size 64 --enable-expert-parallel --speculative-config '{"method":"dspark","num_speculative_tokens":5,"revision":"1bae0e1088993b9b82de08f59de5be5a34e42170","enable_adaptive_verification":true,"use_local_argmax_reduction":true}' --gpu-memory-utilization 0.95 --max-num-batched-tokens 4096 --compilation-config '{"max_cudagraph_capture_size":768,"cudagraph_capture_sizes":[1,2,4,6,8,12,16,24,32,40,48,56,64,72,80,88,96,104,112,120,128,136,144,152,160,168,176,184,192,200,208,216,224,232,240,248,256,272,288,304,320,336,352,368,384,400,416,432,448,464,480,496,512,528,544,560,576,592,608,624,640,656,672,688,704,720,736,752,768]}' --served-model-name deepseek-ai/DeepSeek-V4.1-Flash-1017

Prefill

prompt tok/s by context length
8K context17K481 ms TTFT
16K context17.9K905 ms TTFT
32K context17.9K1,797 ms TTFT
64K context17.9K3,604 ms TTFT
128K context17.5K7,347 ms TTFT
256K context16.8K15,305 ms TTFT
512K context15.4K33,392 ms TTFT
1019K context13.3K76,614 ms TTFT
1,048,512 context13.3K76,981 ms TTFT

Sustained decode

total tok/s · rows are context length, columns are simultaneous requests
Context1248163264128
8K194.483 ms TTFT333.6100 ms TTFT516.8113 ms TTFT773.1353 ms TTFT1,111.0493 ms TTFT1,778.0637 ms TTFT2,627.11,037 ms TTFT3,724.71,801 ms TTFT
16K197.099 ms TTFT322.4117 ms TTFT525.6117 ms TTFT771.2290 ms TTFT1,111.2406 ms TTFT1,753.2725 ms TTFT2,587.81,206 ms TTFTOver capacity
32K187.4146 ms TTFT321.9157 ms TTFT518.9162 ms TTFT785.9401 ms TTFT1,080.4633 ms TTFT1,681.21,119 ms TTFTOver capacityOver capacity
64K198.5219 ms TTFT339.4228 ms TTFT511.8239 ms TTFT742.0617 ms TTFT1,067.81,131 ms TTFTOver capacityOver capacityOver capacity
128K202.1385 ms TTFT324.8393 ms TTFT501.9423 ms TTFT736.51,166 ms TTFTOver capacityOver capacityOver capacityOver capacity
256K181.1725 ms TTFT311.6742 ms TTFT485.61,031 ms TTFTOver capacityOver capacityOver capacityOver capacityOver capacity
512K182.21,542 ms TTFT299.71,549 ms TTFTOver capacityOver capacityOver capacityOver capacityOver capacityOver capacity
1019K161.03,258 ms TTFTOver capacityOver capacityOver capacityOver capacityOver capacityOver capacityOver capacity

Over capacity: not run because this many requests at this context would exceed the server's KV cache.

TTFT is measured after the benchmark warms the server's prompt cache, so it mostly excludes prefill.

All measurements

Latencies are medians. Total tok/s is output across all simultaneous requests.
PhaseContextRequests at onceStatusTotal tok/sPer request tok/sTTFTTime between tokensAverage active requestsCompleted requestsOutput tokensMeasured for
Prefill8K1Valid17K—481 ms——59——
Prefill16K1Valid17.9K—905 ms——30——
Prefill32K1Valid17.9K—1,797 ms——15——
Prefill64K1Valid17.9K—3,604 ms——8——
Prefill128K1Valid17.5K—7,347 ms——4——
Prefill256K1Valid16.8K—15,305 ms——2——
Prefill512K1Valid15.4K—33,392 ms——1——
Prefill1019K1Valid13.3K—76,614 ms——1——
Prefill1,048,5121Valid13.3K—76,981 ms——1——
Sustained decode8K1Valid194.4193.083 ms5.2 ms1.015,83130.0 s
Sustained decode8K2Valid333.6163.8100 ms6.1 ms2.0210,00830.0 s
Sustained decode8K4Valid516.8128.3113 ms7.8 ms4.0415,50430.0 s
Sustained decode8K8Valid773.197.3353 ms10.3 ms8.0023,18630.0 s
Sustained decode8K16Valid1,111.069.6493 ms14.4 ms16.0033,32330.0 s
Sustained decode8K32Valid1,778.056.6637 ms17.7 ms32.0053,29730.0 s
Sustained decode8K64Valid2,627.141.61,037 ms24.0 ms64.0078,80430.0 s
Sustained decode8K128Valid3,724.729.31,801 ms34.2 ms128.00111,69930.0 s
Sustained decode16K1Valid197.0196.899 ms5.1 ms1.015,91030.0 s
Sustained decode16K2Valid322.4163.4117 ms6.1 ms2.029,66930.0 s
Sustained decode16K4Valid525.6130.1117 ms7.7 ms4.0415,76530.0 s
Sustained decode16K8Valid771.297.1290 ms10.3 ms8.0023,12230.0 s
Sustained decode16K16Valid1,111.270.1406 ms14.3 ms16.0033,31430.0 s
Sustained decode16K32Valid1,753.256.1725 ms17.8 ms32.0052,56130.0 s
Sustained decode16K64Valid2,587.840.81,206 ms24.5 ms64.0077,61930.0 s
Sustained decode16K128Over capacity————————
Sustained decode32K1Valid187.4185.0146 ms5.4 ms1.015,62330.0 s
Sustained decode32K2Valid321.9159.3157 ms6.3 ms2.029,65530.0 s
Sustained decode32K4Valid518.9127.9162 ms7.8 ms4.0415,55830.0 s
Sustained decode32K8Valid785.994.4401 ms10.6 ms8.0023,57530.0 s
Sustained decode32K16Valid1,080.467.6633 ms14.8 ms16.0032,40430.0 s
Sustained decode32K32Valid1,681.253.41,119 ms18.7 ms32.0050,40330.0 s
Sustained decode32K64Over capacity————————
Sustained decode32K128Over capacity————————
Sustained decode64K1Valid198.5196.0219 ms5.1 ms1.015,95330.0 s
Sustained decode64K2Valid339.4163.3228 ms6.1 ms2.0210,17930.0 s
Sustained decode64K4Valid511.8127.6239 ms7.8 ms4.0415,33430.0 s
Sustained decode64K8Valid742.092.4617 ms10.8 ms8.0022,25830.0 s
Sustained decode64K16Valid1,067.868.01,131 ms14.7 ms16.0032,03430.0 s
Sustained decode64K32Over capacity————————
Sustained decode64K64Over capacity————————
Sustained decode64K128Over capacity————————
Sustained decode128K1Valid202.1201.3385 ms5.0 ms1.016,06130.0 s
Sustained decode128K2Valid324.8161.5393 ms6.2 ms2.029,74130.0 s
Sustained decode128K4Valid501.9132.8423 ms7.5 ms3.9415,05630.0 s
Sustained decode128K8Valid736.592.71,166 ms10.8 ms8.0022,08630.0 s
Sustained decode128K16Over capacity————————
Sustained decode128K32Over capacity————————
Sustained decode128K64Over capacity————————
Sustained decode128K128Over capacity————————
Sustained decode256K1Valid181.1185.4725 ms5.4 ms1.015,43130.0 s
Sustained decode256K2Valid311.6157.7742 ms6.3 ms2.029,34730.0 s
Sustained decode256K4Valid485.6125.71,031 ms8.0 ms4.0414,56130.0 s
Sustained decode256K8Over capacity————————
Sustained decode256K16Over capacity————————
Sustained decode256K32Over capacity————————
Sustained decode256K64Over capacity————————
Sustained decode256K128Over capacity————————
Sustained decode512K1Valid182.2191.61,542 ms5.2 ms1.015,46730.0 s
Sustained decode512K2Valid299.7153.21,549 ms6.5 ms1.928,99030.0 s
Sustained decode512K4Over capacity————————
Sustained decode512K8Over capacity————————
Sustained decode512K16Over capacity————————
Sustained decode512K32Over capacity————————
Sustained decode512K64Over capacity————————
Sustained decode512K128Over capacity————————
Sustained decode1019K1Valid161.0180.23,258 ms5.6 ms0.914,82930.0 s
Sustained decode1019K2Over capacity————————
Sustained decode1019K4Over capacity————————
Sustained decode1019K8Over capacity————————
Sustained decode1019K16Over capacity————————
Sustained decode1019K32Over capacity————————
Sustained decode1019K64Over capacity————————
Sustained decode1019K128Over capacity————————

Over capacity: not run because this many requests at this context would exceed the server's KV cache.