Open-source · DGX Spark · GB10

What actually runs on a DGX Spark.

SparkBench is a community lab for the NVIDIA GB10. We benchmark every model on real hardware, publish reproducible recipes for coding, agents, and reasoning, and ship the tooling so you can run it on your own box.

Browse the leaderboard Get the tool ↗

SparkBench portal: Inference, Models, and Explore tabs — Portal — switch profiles, browse models, explore HuggingFace

Recipes benched

86.3 t/s @ 4k

Top @ 4k fill

Inference engines

GB10 box

Editor's pick Qwen3.6 35B A3B · MTP 86.3 t/s @ 4k

Leaderboard

Sorted by PBM @ 4k · 32 recipes · Data updated 2026-08-04

Rankings are by throughput (tok/s) on a single GB10 — not by intelligence, coding quality, or SWE benchmarks. We can't wait to add those. Contributions welcome ↗

Measured at 4k context fill (PBM) — decode tok/s after prefilling a fixed token context (perfbench-metrics). Same agent-style workload as bench v2, but comparable across models at 4k / 50k / 100k fills.

#	Model	For	Engine	Params	Curve	Throughput
01	Qwen3.6 35B A3B · MTP nvidia/qwen3.6-35b-a3b NVFP4 MTP · n3	AgentsGeneralReasoning	vLLM	3B / 35B		86.3t/s@ 4k	HF ↗
02	Mellum2 12B MoE Opus Thinking yuxinlu1/mellum2-12b-opus-thinking Q4	CodeReasoning	llama.cpp	2.5B / 12B	—	83.8t/s@ 4k	HF ↗
03	Qwen3-30B-A3B nvidia/qwen3-30b-a3b NVFP4	AgentsGeneralReasoning	vLLM	3B / 30B	—	74.2t/s@ 4k	HF ↗
04	Qwen3.6-35B-A3B NVFP4 (Mia recipe) unsloth/qwen3.6-35b-a3b-nvfp4 NVFP4 MTP · n2	AgentsMultimodal	vLLM	3B / 35B		70.3t/s@ 4k	HF ↗
05	qwen3-coder-next saricles/qwen3-coder-next NVFP4	AgentsCode	vLLM	3B / 80B		58.9t/s@ 4k	HF ↗
06	Qwen3.6 35B unsloth/qwen3.6-35b-a3b-nvfp4-fast Q4	Reasoning	vLLM	3B / 35B	—	—	HF ↗
07	Qwen3.6 35B unsloth/qwen3.6-35b-a3b Q4	AgentsGeneralReasoning	llama.cpp	3B / 35B		46.8t/s@ 4k	HF ↗
08	Ornith 1.0 35B s-batman/ornith-1.0-35b-nvfp4 NVFP4	General	llama.cpp	3B / 35B		46.5t/s@ 4k	HF ↗
09	Qwen3-Coder-30B-A3B-Instruct unsloth/qwen3-coder-30b-a3b-instruct Q4_K_M	AgentsCode	llama.cpp	3B / 30B		42.1t/s@ 4k	HF ↗
10	qwen3-coder-next qwen/qwen3-coder-next	AgentsCode	vLLM	3B / 80B		38.7t/s@ 4k	HF ↗
11	Ornith 1.0 35B deepreinforce-ai/ornith-1.0-35b	General	llama.cpp	3B / 35B		37.8t/s@ 4k	HF ↗
12	qwen-agentworld-35b-a3b qwen/qwen-agentworld-35b-a3b	General	vLLM	3B / 35B		29.6t/s@ 4k	HF ↗
13	Laguna S 2.1 poolside/laguna-s-2.1 NVFP4 DFlash · n3	AgentsCode	vLLM	8B / 118B		29.1t/s@ 4k	HF ↗
14	qwen3.6-27b-aeon-ultimate-uncensored aeon-7/qwen3.6-27b-aeon-ultimate-uncensored NVFP4 DFlash · n10	General	vLLM	27B		26.0t/s@ 4k	HF ↗
15	glm-4.7-flash-30b-a3b-q4_k_m-5gb inference-snaps/glm-4.7-flash-30b-a3b-q4_k_m-5gb Q4	General	llama.cpp	3B / 30B	—	—	HF ↗
16	Gemma-4-26B-A4B-IT google/gemma-4-26b-a4b-it	GeneralMultimodalReasoning	vLLM	4B / 26B		23.0t/s@ 4k	HF ↗
17	Laguna XS 2.1 poolside/laguna-xs-2.1 DFlash · n15	AgentsCode	vLLM	3B / 33B		21.4t/s@ 4k	HF ↗
18	Step-3.7-Flash stepfun-ai/step-3.7-flash IQ4_XS	AgentsMultimodal	llama.cpp	11B / 198B		20.5t/s@ 4k	HF ↗
19	Gemma-4-12B-IT google/gemma-4-12b-it Q4_K_M	GeneralMultimodal	llama.cpp	12B		19.8t/s@ 4k	HF ↗
20	DeepSeek V4 Flash (DwarfStar) antirez/deepseek-v4-flash	AgentsCodeReasoning	ds4	13B / 180B	—	—	HF ↗
21	Gemma 4 12B Coder (Fable5×Composer2.5) yuxinlu1/gemma-4-12b-coder-fable5-composer2.5-v1 Q4	Code	llama.cpp	12B		19.1t/s@ 4k	HF ↗
22	Gemma 4 12B Opus Reasoning yuxinlu1/gemma-4-12b-opus-reasoning Q4	Reasoning	llama.cpp	12B	—	19.1t/s@ 4k	HF ↗
23	Gemma 4 12B Agentic v2 (Fable5×Composer2.5) yuxinlu1/gemma-4-12b-agentic-v2 Q4	AgentsCode	llama.cpp	12B	—	18.9t/s@ 4k	HF ↗
24	Qwen3.6-27B qwen/qwen3.6-27b FP8 DFlash · n10	AgentsGeneralReasoning	vLLM	27B		17.3t/s@ 4k	HF ↗
25	qwythos-9b-claude-mythos-5-1m empero-ai/qwythos-9b-claude-mythos-5-1m	General	vLLM	9B		12.6t/s@ 4k	HF ↗
26	Qwen3 6 27B nvidia/qwen3.6-27b NVFP4	General	vLLM	27B		12.1t/s@ 4k	HF ↗
27	qwen3.6-27b-aeon-ultimate-uncensored-text-nvfp4-mtp-xs aeon-7/qwen3.6-27b-aeon-ultimate-uncensored-text-nvfp4-mtp-xs NVFP4 MTP	General	vLLM	27B	—	—	HF ↗
28	thinkingcap-qwen3.6-27b sakamakismile/thinkingcap-qwen3.6-27b	General	vLLM	27B	—	—	private
29	Qwen3.6-27B rdtand/qwen3.6-27b PrismaQuant	GeneralReasoning	vLLM	27B		10.6t/s@ 4k	HF ↗
30	Qwen3.6-27B (unsloth) unsloth/qwen3.6-27b NVFP4	AgentsGeneralReasoning	vLLM	27B		9.4t/s@ 4k	HF ↗
31	Qwen3.6-27B kaitchup/qwen3.6-27b	GeneralReasoning	llama.cpp	27B		9.1t/s@ 4k	HF ↗
32	thinkingcap-qwen3.6-27b protolabsai/thinkingcap-qwen3.6-27b Q5_K_M	General	llama.cpp	27B	—	—	private

Recipes by task

Pick by what you're building

General

Agents

Reasoning

Code

Multimodal

Run it on your own Spark

Same tool that generates this leaderboard

SparkBench is the operator tool behind this site. Portal, model inventory, three inference engines (vLLM, llama.cpp, ds4), reproducible bench v2, and an OpenAI gateway — one CLI on your GB10.

One bootstrap command — clone, host env, portal, APIs, CLI.
Three engines — eugr, llama.cpp, ds4. One GPU at a time.
Real recipes — auto-scaffolded from weights; golden map in git.
Your data, your box — nothing leaves the LAN unless you ship it here.

        install
        shell
      

# One command — core stack (no GPU engine yet)
curl -fsSL https://raw.githubusercontent.com/shawnmarck/sparkbench/main/scripts/bootstrap-sparkbench.sh | sudo bash

# Then pick an engine + gateway
sudo bash install/spark-install engine eugr
sudo bash install/spark-install gateway

# Switch, bench, serve
spark inference list
spark inference up qwen36-nvfp4
spark inference bench