Spark Bench

DGX Spark benchmarks

Model catalog

Filter by hardware, serving engine, and evaluation results. Browse by model or sort across configurations by speed or date.

32 configurations across 15 models

Decode sweep · ~256 input / 512 output tokens per request

DeepSeek V4.1 Flash 1017 EXL3-EAQ

2 configurations

2× DGX SparkTensorFold

DeepSeek V4.1 Flash 1017 EXL3-EAQ 2.9 bpw

TensorFold native, 2 Sparks TP=2, 262K, 4 streams, recipe 27741e17

Single-stream decode
67.50output tok/s
TTFT p95 · c1
0.51 s
Peak decode
125.87output tok/s
c16
Sustained decode
108.73output tok/s
c8 · 20 min soak
Prefill
2,008.07input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval incompleteSafety gate passed
View benchmarkView recipe
2× DGX SparkTensorFold

DeepSeek V4.1 Flash 1017 EXL3-EAQ 2.9 bpw

TensorFold native, 2 Sparks TP=2, 262K, 4 streams

Single-stream decode
65.62output tok/s
TTFT p95 · c1
0.48 s
Peak decode
123.18output tok/s
c16
Sustained decode
103.17output tok/s
c8 · 20 min soak
Prefill
1,719.72input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval incompleteSafety gate passed
View benchmarkView recipe

GLM-5.3 Flash

7 configurations

2× DGX SparkTensorFold

GLM-5.3 Flash EXL3 4 bpw

TensorFold 1.3.0, DFlash2, P4, recipe a691916c

Single-stream decode
41.85output tok/s
TTFT p95 · c1
0.51 s
Peak decode
77.02output tok/s
c8
Sustained decode
65.01output tok/s
c8 · 20 min soak
Prefill
1,170.32input tok/s
32K prompt · c1

Context

Configured
1,048,576 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe
2× DGX SparkTensorFold

GLM-5.3 Flash EXL3 4 bpw

TensorFold 1.1.9, DFlash2, recipe 13f2464

Single-stream decode
42.67output tok/s
TTFT p95 · c1
0.37 s
Peak decode
121.04output tok/s
c16
Sustained decode
62.29output tok/s
c8 · 20 min soak
Prefill
1,569.79input tok/s
32K prompt · c1

Context

Configured
1,048,576 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe
2× DGX SparkTensorFold

GLM-5.3 Flash EXL3 4 bpw

TensorFold v0.6.0, DFlash2, recipe 3fd60dfc

Single-stream decode
54.53output tok/s
TTFT p95 · c1
0.44 s
Peak decode
93.20output tok/s
c4
Sustained decode
80.18output tok/s
c8 · 20 min soak
Prefill
1,744.96input tok/s
32K prompt · c1

Context

Configured
1,048,576 tokens
Longest tested prompt
131,072 target tokens
Validated retrieval
Unavailable

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe
2× DGX SparkvLLM

GLM-5.3 Flash NVFP4

vLLM, low reasoning

Single-stream decode
14.20output tok/s
TTFT p95 · c1
0.45 s
Peak decode
66.18output tok/s
c16
Sustained decode
60.90output tok/s
c8 · 20 min soak
Prefill
1,096.57input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe
2× DGX SparkTensorFold

GLM-5.3 Flash EXL3/TR3 4 bpw

TensorFold v0.6.0

Single-stream decode
51.49output tok/s
TTFT p95 · c1
0.41 s
Peak decode
96.72output tok/s
c12
Sustained decode
84.90output tok/s
c8 · 20 min soak
Prefill
1,662.87input tok/s
32K prompt · c1

Context

Configured
1,048,576 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe
2× DGX SparkvLLM

GLM-5.3 Flash EXL3/TR3 4 bpw

refreshed recipe

Single-stream decode
13.96output tok/s
TTFT p95 · c1
1.27 s
Peak decode
30.68output tok/s
c4
Sustained decode
21.88output tok/s
c8 · 20 min soak
Prefill
781.30input tok/s
32K prompt · c1

Context

Configured
850,000 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe
2× DGX SparkvLLM

GLM-5.3-Flash-EXL3

Single-stream decode
35.13output tok/s
TTFT p95 · c1
0.95 s
Peak decode
79.51output tok/s
c16
Sustained decode
38.53output tok/s
c8 · 20 min soak
Prefill
833.16input tok/s
32K prompt · c1

Context

Configured
1,000,000 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe

Qwen3.8 27B

3 configurations

1× DGX SparkTensorFold

Qwen3.8 27B MLX 4-bit

TensorFold, native 262K, thinking off

Single-stream decode
50.25output tok/s
TTFT p95 · c1
0.44 s
Peak decode
201.49output tok/s
c16
Sustained decode
160.59output tok/s
c8 · 20 min soak
Prefill
1,478.42input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval incompleteSafety gate failed
View benchmarkView recipe
1× DGX SparkSGLang

Qwen3.8 27B NVFP4

SGLang

Single-stream decode
31.83output tok/s
TTFT p95 · c1
0.18 s
Peak decode
187.63output tok/s
c16
Sustained decode
157.80output tok/s
c8 · 20 min soak
Prefill
1,402.94input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe
1× DGX SparkSGLang

Qwen3.8 27B Uncensored FP8

SGLang

Single-stream decode
18.60output tok/s
TTFT p95 · c1
0.28 s
Peak decode
72.86output tok/s
c4
Sustained decode
63.89output tok/s
c8 · 20 min soak
Prefill
710.55input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe

Qwen3.8-Flash-Next NVFP4

9 configurations

2× DGX SparkvLLM 0.34.0

Qwen3.8-Flash-Next NVIDIA NVFP4

vLLM 0.34.0, BF16 KV

Single-stream decode
44.35output tok/s
TTFT p95 · c1
0.36 s
Peak decode
103.83output tok/s
c4
Sustained decode
140.54output tok/s
c8 · 20 min soak
Prefill
2,311.42input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe
1× DGX SparkTensorFold

Qwen3.8-Flash-Next FP4-Local QAD

TensorFold v1.1.4, 1 Spark, 262K, recipe 20e15e4

Single-stream decode
41.54output tok/s
TTFT p95 · c1
0.35 s
Peak decode
111.73output tok/s
c4
Sustained decode
77.79output tok/s
c8 · 20 min soak
Prefill
2,059.50input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe
1× DGX SparkTensorFold

Qwen3.8-Flash-Next FP4-Local QAD

TensorFold, 1 Spark, native 262K, thinking off

Single-stream decode
34.29output tok/s
TTFT p95 · c1
0.48 s
Peak decode
100.43output tok/s
c12
Sustained decode
85.72output tok/s
c8 · 20 min soak
Prefill
1,962.37input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval incompleteSafety gate passed
View benchmarkView recipe
2× DGX SparkTensorFold

Qwen3.8-Flash-Next NVIDIA NVFP4

TensorFold, 1M YaRN

Single-stream decode
83.45output tok/s
TTFT p95 · c1
0.23 s
Peak decode
417.63output tok/s
c16
Sustained decode
432.90output tok/s
c8 · 20 min soak
Prefill
2,673.36input tok/s
32K prompt · c1

Context

Configured
1,048,576 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval incompleteSafety gate passed

External benchmark pilots

View benchmarkView recipe
2× DGX SparkvLLM

Qwen3.8-Flash-Next NVIDIA NVFP4

Mia vLLM, FP8 KV, 2200 MHz

Single-stream decode
44.62output tok/s
TTFT p95 · c1
0.42 s
Peak decode
205.06output tok/s
c12
Sustained decode
196.70output tok/s
c8 · 20 min soak
Prefill
2,527.51input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe
2× DGX SparkvLLM

Qwen3.8-Flash-Next NVIDIA NVFP4

Mia vLLM, FP8 KV

Single-stream decode
47.34output tok/s
TTFT p95 · c1
0.40 s
Peak decode
207.77output tok/s
c16
Sustained decode
197.42output tok/s
c8 · 20 min soak
Prefill
2,615.32input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe
2× DGX SparkvLLM

Qwen3.8-Flash-Next-NVFP4

Mia vLLM, 2200 MHz

Single-stream decode
47.51output tok/s
TTFT p95 · c1
0.36 s
Peak decode
213.72output tok/s
c8
Sustained decode
201.58output tok/s
c8 · 20 min soak
Prefill
2,361.85input tok/s
32K prompt · c1

Context

Configured
1,000,000 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe
2× DGX SparkvLLM

Qwen3.8-Flash-Next-NVFP4

Mia vLLM

Single-stream decode
48.93output tok/s
TTFT p95 · c1
0.43 s
Peak decode
209.12output tok/s
c12
Sustained decode
197.71output tok/s
c8 · 20 min soak
Prefill
2,399.48input tok/s
32K prompt · c1

Context

Configured
1,000,000 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe
2× DGX SparkSGLang

Qwen3.8-Flash-Next-NVFP4

Single-stream decode
38.43output tok/s
TTFT p95 · c1
0.41 s
Peak decode
215.06output tok/s
c8
Sustained decode
157.61output tok/s
c8 · 20 min soak
Prefill
1,551.13input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe

DeepSeek V4 Flash 0731

1 configuration

1× DGX SparkvLLM + SparkInfer

DeepSeek V4 Flash 0731 EXL3/SparkInfer

Single-stream decode
29.62output tok/s
TTFT p95 · c1
0.54 s
Peak decode
31.44output tok/s
c4
Sustained decode
26.89output tok/s
c8 · 20 min soak
Prefill
1,046.89input tok/s
32K prompt · c1

Context

Configured
384,000 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe

DeepSeek V4 Flash DSpark

1 configuration

2× DGX SparkvLLM (Anemll DSpark)

DeepSeek V4 Flash DSpark

Anemll DSpark vLLM 0.1.1, original July deployment, 2 Sparks TP=2, 1M, thinking off

Single-stream decode
42.12output tok/s
TTFT p95 · c1
4.12 s
Peak decode
104.82output tok/s
c16
Sustained decode
111.10output tok/s
c8 · 20 min soak
Prefill
1,304.90input tok/s
32K prompt · c1

Context

Configured
1,048,576 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate passed
View benchmarkView recipe

DeepSeek V4 Flash Vision-Exp

1 configuration

2× DGX SparkvLLM (Anemll DSpark 0.1.1)

DeepSeek V4 Flash Vision-Exp

Anemll vLLM DSpark, TP=2, 1M, thinking off

Single-stream decode
25.93output tok/s
TTFT p95 · c1
0.69 s
Peak decode
79.91output tok/s
c8
Sustained decode
84.53output tok/s
c8 · 20 min soak
Prefill
1,334.30input tok/s
32K prompt · c1

Context

Configured
1,048,576 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe

DeepSeek V4.1 Flash EXL3

1 configuration

2× DGX SparkvLLM

DeepSeek V4.1 Flash EXL3 2.9 bpw

Single-stream decode
27.42output tok/s
TTFT p95 · c1
0.85 s
Peak decode
44.49output tok/s
c2
Sustained decode
35.31output tok/s
c8 · 20 min soak
Prefill
559.95input tok/s
32K prompt · c1

Context

Configured
600,000 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval unavailable
View benchmarkView recipe

MiniCPM-V 4.7 35B-A3B

1 configuration

1× DGX SparkExLlamaV3

MiniCPM-V 4.7 35B-A3B EXL3 6.08bpw

ExLlamaV3, native 262K, thinking off

Single-stream decode
47.74output tok/s
TTFT p95 · c1
2.74 s
Peak decode
52.74output tok/s
c16
Sustained decode
51.35output tok/s
c8 · 20 min soak
Prefill
2,344.87input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval incompleteSafety gate failed
View benchmarkView recipe

NousResearch Arete-27B

1 configuration

1× DGX SparkSGLang

NousResearch Arete-27B

SGLang, FP8 KV, 1 Spark, 262K

Single-stream decode
4.29output tok/s
TTFT p95 · c1
0.61 s
Peak decode
32.65output tok/s
c16
Sustained decode
31.28output tok/s
c8 · 20 min soak
Prefill
719.69input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate passed
View benchmarkView model card

NVIDIA Nemotron 3.5 Bolt 74B-A9B

1 configuration

1× DGX SparkTensorFold

Nemotron 3.5 Bolt 74B-A9B 24GB

TensorFold, native 262K

Single-stream decode
79.23output tok/s
TTFT p95 · c1
0.44 s
Peak decode
234.45output tok/s
c16
Sustained decode
145.52output tok/s
c8 · 20 min soak
Prefill
1,421.49input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe

NVIDIA Nemotron 3.5 Super VL 120B-A12B

1 configuration

1× DGX SparkvLLM

Nemotron 3.5 Super VL 120B-A12B NVFP4

vLLM + DFlash LiLiCorr, 1 Spark, 1M, thinking off

Single-stream decode
34.15output tok/s
TTFT p95 · c1
0.58 s
Peak decode
34.95output tok/s
c12
Sustained decode
33.90output tok/s
c8 · 20 min soak
Prefill
1,934.07input tok/s
32K prompt · c1

Context

Configured
1,048,576 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe

Qwen3.6-35B-A3B NVFP4 (Unsloth)

1 configuration

1× DGX SparkvLLM

Qwen3.6-35B-A3B NVFP4

vLLM GB10 + MiaAI-Lab patch, MTP 2, 1 Spark, 262K

Single-stream decode
77.46output tok/s
TTFT p95 · c1
0.20 s
Peak decode
429.21output tok/s
c16
Sustained decode
286.66output tok/s
c8 · 20 min soak
Prefill
3,711.67input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
131,072 target tokens
Validated retrieval
Unavailable

Tested

Tool eval completeSafety gate failed
View benchmarkView recipe

Recursive Flint 40B-A2.5B NVFP4 GGUF

1 configuration

1× DGX Sparkllama.cpp

Recursive Flint NVFP4 GGUF

llama.cpp b11856, -np 1, 1 Spark, 262K

Single-stream decode
81.73output tok/s
TTFT p95 · c1
0.32 s
Peak decode
83.40output tok/s
c2
Sustained decode
74.62output tok/s
c8 · 20 min soak
Prefill
1,848.59input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate passed
View benchmarkView model card

Thinking Machines Inkling 2 Mini

1 configuration

1× DGX SparkvLLM + DSpark7

Inkling 2 Mini NVFP4

vLLM, DSpark7, 1 Spark, 262K

Single-stream decode
41.59output tok/s
TTFT p95 · c1
0.28 s
Peak decode
229.47output tok/s
c16
Sustained decode
145.04output tok/s
c8 · 20 min soak
Prefill
2,363.60input tok/s
32K prompt · c1

Context

Configured
262,144 tokens
Longest tested prompt
245,760 target tokens
Validated retrieval
≈240K passkey

Tested

Tool eval completeSafety gate failed

External benchmark pilots

Install model

Copy this prompt and paste it into an AI assistant with access to your DGX Spark.