Spark Bench

DGX Spark benchmarks

Model catalog

Filter by hardware and serving engine. Configurations are grouped by model so you can compare runs directly.

30 configurations across 14 models

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

DeepSeek V4.1 Flash 1017 EXL3-EAQ

2 configurations

GLM-5.3 Flash

7 configurations

2× DGX Spark

GLM-5.3 Flash EXL3 4 bpw (TensorFold 1.1.9, DFlash2, recipe 13f2464)

TensorFold

Single-stream decode
42.67output tok/s · c1
Sustained throughput
62.29output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 1,048,576 tokens Validated ≈240K passkey

Tool eval completeSafety gate failed
2× DGX Spark

GLM-5.3 Flash NVFP4 (vLLM, low reasoning)

vLLM

Single-stream decode
14.20output tok/s · c1
Sustained throughput
60.90output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval unavailable
2× DGX Spark

GLM-5.3 Flash EXL3/TR3 4 bpw (TensorFold v0.6.0)

TensorFold

Single-stream decode
51.49output tok/s · c1
Sustained throughput
84.90output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 1,048,576 tokens Validated ≈240K passkey

Tool eval unavailable
2× DGX Spark

GLM-5.3 Flash EXL3/TR3 4 bpw (refreshed recipe)

vLLM

Single-stream decode
13.96output tok/s · c1
Sustained throughput
21.88output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 850,000 tokens Validated ≈240K passkey

Tool eval unavailable
2× DGX Spark

GLM-5.3-Flash-EXL3

vLLM

Single-stream decode
35.13output tok/s · c1
Sustained throughput
38.53output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 1,000,000 tokens Validated ≈240K passkey

Tool eval unavailable

Qwen3.8 27B

3 configurations

1× DGX Spark

Qwen3.8 27B MLX 4-bit (TensorFold, native 262K, thinking off)

TensorFold

Single-stream decode
50.25output tok/s · c1
Sustained throughput
160.59output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval incompleteSafety gate failed
1× DGX Spark

Qwen3.8 27B NVFP4 (SGLang)

SGLang

Single-stream decode
31.83output tok/s · c1
Sustained throughput
157.80output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval unavailable
1× DGX Spark

Qwen3.8 27B Uncensored FP8 (SGLang)

SGLang

Single-stream decode
18.60output tok/s · c1
Sustained throughput
63.89output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval unavailable

Qwen3.8-Flash-Next NVFP4

8 configurations

2× DGX Spark

Qwen3.8-Flash-Next NVIDIA NVFP4 (TensorFold, 1M YaRN)

TensorFold

Single-stream decode
83.45output tok/s · c1
Sustained throughput
432.90output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 1,048,576 tokens Validated ≈240K passkey

Tool eval incompleteSafety gate passed
2× DGX Spark

Qwen3.8-Flash-Next NVIDIA NVFP4 (Mia vLLM, FP8 KV, 2200 MHz)

vLLM

Single-stream decode
44.62output tok/s · c1
Sustained throughput
196.70output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval unavailable
2× DGX Spark

Qwen3.8-Flash-Next NVIDIA NVFP4 (Mia vLLM, FP8 KV)

vLLM

Single-stream decode
47.34output tok/s · c1
Sustained throughput
197.42output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval unavailable
2× DGX Spark

Qwen3.8-Flash-Next-NVFP4 (Mia vLLM, 2200 MHz)

vLLM

Single-stream decode
47.51output tok/s · c1
Sustained throughput
201.58output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 1,000,000 tokens Validated ≈240K passkey

Tool eval unavailable
2× DGX Spark

Qwen3.8-Flash-Next-NVFP4 (Mia vLLM)

vLLM

Single-stream decode
48.93output tok/s · c1
Sustained throughput
197.71output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 1,000,000 tokens Validated ≈240K passkey

Tool eval unavailable
2× DGX Spark

Qwen3.8-Flash-Next-NVFP4

SGLang

Single-stream decode
38.43output tok/s · c1
Sustained throughput
157.61output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval unavailable

DeepSeek V4 Flash 0731

1 configuration

1× DGX Spark

DeepSeek V4 Flash 0731 EXL3/SparkInfer

vLLM + SparkInfer

Single-stream decode
29.62output tok/s · c1
Sustained throughput
26.89output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 384,000 tokens Validated ≈240K passkey

Tool eval unavailable

DeepSeek V4 Flash DSpark

1 configuration

DeepSeek V4 Flash Vision-Exp

1 configuration

DeepSeek V4.1 Flash EXL3

1 configuration

2× DGX Spark

DeepSeek V4.1 Flash EXL3 2.9 bpw

vLLM

Single-stream decode
27.42output tok/s · c1
Sustained throughput
35.31output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 600,000 tokens Validated ≈240K passkey

Tool eval unavailable

MiniCPM-V 4.7 35B-A3B

1 configuration

NousResearch Arete-27B

1 configuration

1× DGX Spark

NousResearch Arete-27B (SGLang, FP8 KV, 1 Spark, 262K)

SGLang

Single-stream decode
4.29output tok/s · c1
Sustained throughput
31.28output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval completeSafety gate passed

NVIDIA Nemotron 3.5 Bolt 74B-A9B

1 configuration

1× DGX Spark

Nemotron 3.5 Bolt 74B-A9B 24GB (TensorFold, native 262K)

TensorFold

Single-stream decode
79.23output tok/s · c1
Sustained throughput
145.52output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval completeSafety gate failed

NVIDIA Nemotron 3.5 Super VL 120B-A12B

1 configuration

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

1 configuration

Recursive Flint 40B-A2.5B NVFP4 GGUF

1 configuration

1× DGX Spark

Recursive Flint NVFP4 GGUF (llama.cpp b11856, -np 1, 1 Spark, 262K)

llama.cpp

Single-stream decode
81.73output tok/s · c1
Sustained throughput
74.62output tok/s · c8 · 20 min
Longest tested prompt
245,760target input tokens · c1

Configured 262,144 tokens Validated ≈240K passkey

Tool eval completeSafety gate passed