Start here
- NVIDIA Spark Playbooks — Step-by-step AI workloads published for DGX Spark.
- DGX Spark setup guide — Practical single-node setup, recommended settings and troubleshooting.
Videos and walkthroughs
- Eight-Spark GLM-5.2 cluster at 400K context — YourAvgDev builds an eight-node DGX Spark cluster and runs GLM-5.2 with a 400K context window.
- Bare-metal vLLM cluster with Qwen3.5 122B — YourAvgDev serves Qwen3.5 122B-A10B-FP8 at full context across a DGX Spark cluster.
- Three-node DGX Spark cluster setup — Justin Garrison walks through configuring a three-node Spark cluster.
- Eight DGX Sparks with 1TB of unified memory — Alex Ziskind connects eight Sparks with high-speed networking and tests large local models.
- GB10 thermals, power limits and storage — Alex Ziskind stress-tests DGX Spark and three GB10 alternatives to investigate throttling claims.
- DGX Spark-class 4TB storage upgrade — Alex Ziskind upgrades a GB10 system and examines its storage and value tradeoffs.
- Eight-node GB10 local AI cluster — ServeTheHome builds a mixed-vendor GB10 cluster with 1TB of aggregate memory and 400GbE networking.
Recipe collections and maintainers
Use these profile and collection links to find each maintainer's current recipes. This stays useful as model versions and repository names change.
- Club DGX Spark notebooks — Community notebooks for multi-node serving, vision and performance experiments.
- Inference Atlas — Browse engines and deployment targets, including Spark image-generation setups.
DeepSeek recipe maintainers
GLM recipe maintainers
- MiaAI-Lab — GLM EXL3 deployments for two-to-four-Spark clusters.
- tonyd2wild — GLM multi-node recipes, memory studies and ComfyUI fleet tools.
- alicankiraz1 — GLM NVFP4 and DFlash2 multi-node recipes.
- Reederey87 — GLM EXL3 releases with adaptive decode and memory tuning.
- NNNtrance — Measured GLM recipes with loader, kernel and NCCL fixes.
- beastllama — SGLang and DFlash2 recipes for GLM on two Sparks.
- alexellis — Switchless RoCE cluster recipes and practical deployment notes.
- jnardiello — FP8 multi-node recipes and patched NCCL ring work.
- joesinvestments — Source-form GLM runtime kits, launchers and measured logs.
Qwen, Ling and Kimi recipe maintainers
- vcruz305 — Single-Spark Qwen EXL3 and vLLM recipes.
- DJLougen — Single-Spark Qwen serving projects.
- r0b0tlab — Ling multimodal quantization, runtime containers and evidence.
- sojufx — Ling deployment recipes for DGX Spark.
- ciprianveg — GB10 vLLM work for Kimi, DSpark, B12X and GLM.
Benchmarks and comparisons
- Spark Arena — Browse community benchmark records for DGX Spark models.
- AEON Bench — Community benchmark board for local inference systems.
- Qwen and GLM on DGX Spark — Japanese comparison of four single-node model configurations.
- DeepSeek V4.1 TP2 versus TP4 benchmark — Frozen trials, offline scoring reproduction and sanitized evidence.
Image and multimodal
- Qwen Image 2.1 on one Spark — Day-one ComfyUI setup and validation notes for a single Spark.
- Qwen Image Spark engine — Inference Atlas engine page for running Qwen Image on Spark.
- DiffusionGemma on Spark — Container recipe for NVFP4 structured decisions on DGX Spark.
- ComfyUI Studio Remote — Lightweight web control for a fleet of ComfyUI image and video workers.
Networking and clusters
- NVIDIA ConnectX-7 clustering guide — Official port, cable and cluster-connection guidance.
- NCCL tests — Check communication and collective performance between GPUs.
- MCDMA user guide — Practical guide to multi-computer DMA and high-speed cluster links.
- Dual-Spark vLLM Docker — Docker configuration for serving models across two DGX Sparks.
Hardware, power and cooling
- DGX cluster thermal-management guide — Cooling layout, monitoring and operating guidance for a multi-Spark cluster.
- Guarded power recovery for eight Sparks — Evidence-led rack power design with conservative recovery gates.
- GB10 clock-cap study — Measured temperature, power and throughput tradeoffs from capping GPU clocks.
- SparkStack source files — Stackable 3D-printable rack and panels for DGX Spark hardware.
- SparkStack model on MakerWorld — Ready-to-download files for the stackable desktop rack.
Engines, kernels and optimization
- vLLM — Model serving with batching and parallelism.
- SGLang — Serving for language and multimodal models.
- SGLang on DGX Spark — NVIDIA's Spark-specific installation and inference cookbook.
- Atlas Inference for DGX Spark — Open-source Rust and CUDA inference engine tuned and verified for GB10.
- llama.cpp — Local inference with GGUF models.
- B12X — Local inference kernels and tooling used by several GB10 recipes.
- AEON vLLM Ultimate container — Prebuilt DGX Spark container from the AEON project.
- NVIDIA Model Optimizer — Official quantization, pruning and speculative-decoding toolkit.
- Nsight Python — Python interface for profiling GPU kernels with NVIDIA Nsight tools.
- Auto GPU Kernel — Winning agent-generated DeepSeek sparse-attention kernel project.
Community software
- FrostByte App — Downloads, documentation and issue tracking for the FrostByte local-AI app.
Nix and NixOS
- Graham33's DGX Spark project — NixOS hardware support and Nix playbooks for DGX OS.
- nix.dev — Learn Nix, from your first shell to reproducible builds.
